dynamics, causation, duration in the predicate-argument ... · universit e de gen eve facult e des...

Universite de Geneve

Faculte des Lettres

Doctoral Dissertation

Dynamics, causation, duration in the

predicate-argument structure of verbs:

A computational approach based on

parallel corpora

Tanja Samardzic

June 25, 2014

Supervisor: Prof. Paola Merlo

Abstract

This dissertation addresses systematic variation in the use of verbs where two syntacti-

cally different sentences are used to express the same event, such as the alternations in

the use of decide, break, and push shown in (0.1-0.3). We study the frequency distribu-

tion of the syntactic alternants showing that the distributional patterns originate in the

meaning of the verbs.

(0.1) a. Mary took/made a decision. b. Mary decided (something).

(0.2) a. Adam broke the laptop. b. The laptop broke.

(0.3) a. John pushed the cart. b. John pushed the cart for some time.

Both intra-linguistic and cross-linguistic variation in morphological and syntactic re-

alisations of semantically equivalent items are taken into account by analysing data

extracted from parallel corpora. The dissertation includes three case studies: light verb

constructions (0.1) in English and German, lexical causatives (0.2) also in English and

German, and verb aspect classes (0.3) in English and Serbian.

The core question regarding light verb constructions is whether the verbs such as take

and make, when used in expressions such as (0.1a), turn into functional words losing

their lexical meaning. Arguments for both a positive and a negative answer have been

put forward in the literature. The results of our study suggest that light verbs keep at

least force-dynamic semantics of their lexical counterparts: the inward dynamics in verbs

such as take and the outward dynamics in verbs such as make. The inward dynamics

results in a cross-linguistic preference for compact grammatical forms (single verbs) and

the outward dynamics results in a preference for analytical forms (constructions).

3

The study on lexical causatives (0.2) addresses the question of why some verbs in some

languages do not alternate while their counterparts in other languages do. The results

of the study suggest that the property which underlies the variation is the likelihood of

external causation. Events described by the alternating verbs are distributed on a scale

of increasing likelihood for an external causer to occur. The verbs which alternate in

some but not in other languages are those verbs which describe events on the two ex-

tremes of the scale. The preference for one alternant is so strong in these verbs that the

other alternant rarely occurs, which is why it is not attested in some languages. There

are two ways in which the likelihood of external causation can be empirically assessed:

a) by observing the typological distribution of causative vs. anticausative morpholog-

ical marking across a wide range of languages and b) by the frequency distribution of

transitive vs. intransitive uses of the alternating verbs in a corpus of a single language.

Our study shows that these two measures are correlated. By applying the corpus-based

measure, the position on the scale of likelihood of external causation can be determined

automatically for a wide range of verbs.

The subject of the third case study is the relationship between two temporal properties

encoded by the grammatical category of verb aspect: event duration and temporal

boundedness. The study shows that these two properties interact in a complex but

predictable way giving rise to the observed variation in morphosyntactic realisations of

verbs. English native speakers’ intuitions about possible duration of events described by

verbs (short vs. long) are predicted from the patterns of formal aspect marking in the

equivalent Serbian verbs. The accuracy of the prediction based on the bi-lingual model

is superior to the best performing monolingual model.

One of the main contributions of the dissertation is a novel experimental methodology,

which relies on automatic processing of parallel corpora and statistical inference. The

three properties of the events described by verbs (dynamics orientation, the likelihood

of external causation, duration) are empirically induced on the basis of the observations

automatically extracted from large parallel corpora (containing up to over a million

sentences per language), which are automatically parsed and word-aligned. The gener-

alisations are learned from the extracted data automatically using statistical inferences

and machine learning techniques. The accuracy of the predictions made on the basis of

the generalisations is assessed experimentally on an independent set of test instances.

4

Resume

Cette these porte sur la variation systematique dans l’usage des verbes ou deux phrases,

differentes par rapport a leurs structures syntactiques, peuvent etre utilisees pour ex-

primer le meme evenement. La variation concernee est montree dans les exemples (0.4-

0.6). Nous etudions la distribution des frequences des alternants syntactiques en mon-

trant que la source des patterns distributionnels est dans le contenu semantique des

verbes.

(0.4) a. MaryMarie

took/madepris/fait

aune

decision.decision

Marie a pris une decision.

b. MaryMarie

decideddecide

(something).(quelque chose)

Marie a decide (quelque chose)

(0.5) a. AdamAdam

brokecasse

thele

laptop.ordinateur

Adam a casse l’ordinateur.

b. Thele

laptopordinateur

broke.casse

L’ordinateur c’est casse.

(0.6) a. JohnJean

pushedpousse

thele

cart.chariot

Jean a pousse le chariot.

5

b. JohnJean

pushedpousse

thele

cartchariot

forpour

somequelque

time.temps

Jean poussait le chariot pendent quelque temps.

La variation intra-linguistique ainsi que la variation a travers des langues concernant

les realisations morphologiques et syntactiques des items semantiquement equivalents

sont prises en compte. Ceci est effectue par une analyse des donnees extraites de corpus

paralleles. La these contient trois etudes de cas: constructions a verbes legers (0.4) en

anglais et allemand, les verbes causatifs lexicaux (0.5), egalement en anglais et allemand,

et les classes d’aspect verbal (0.6) en anglais et serbe.

La question centrale par rapport aux constructions a verbes legers est de savoir si les

verbes comme take et make utilises dans des expressions comme (0.4a) deviennent

des mots fonctionnels perdant donc entierement leur contenu lexical. Des arguments

en faveur des deux reponses, positive et negative, ont ete cites dans la litterature.

Les resultats de notre etude suggerent que les verbes legers maintiennent au moins la

semantique de dynamique de force appartenant au contenu des verbes lexicaux equivalents:

La dynamique orientee vers l’agent de l’evenement (a l’interieur) des verbes comme

take et la dynamique orientee vers d’autres participants dans l’evenement (a l’exterieur)

des verbes comme make. La dynamique orientee vers l’interieur a pour consequence

une preference pour des realisations compactes (des verbes individuelles) a travers des

langues, tandis que la dynamique orientee vers l’exterieur a pour consequence une

preference pour des formes analytiques (des constructions).

L’etude des verbes causatifs lexicaux (0.5) porte sur la variation a travers des langues

concernant la participation de ces verbes dans l’alternance causative: Pourquoi certains

verbes dans certaines langues n’entrent pas dans l’alternance causative tandis que leurs

verbes correspondants dans d’autres langues le font? Les resultats de l’etude suggerent

que la caracteristique semantique qui est a la source de la variation est la probabilite

de la causalite externe de l’evenement decrit par un verbe. Les evenements decrits par

les verbes causatifs lexicaux sont places au long d’une echelle de probabilite croissante

de la causalite externe. Les verbes qui entrent dans l’alternance dans une langue, mais

ne le font pas dans d’autres langues, sont les verbes decrivant des evenements qui se

trouvent aux deux extremites de l’echelle. Ces verbes ont une preference pour l’un

6

des deux alternants si forte que l’autre alternant n’apparaıt que rarement. Ceci est la

raison pour laquelle un de deux alternants n’est pas observe dans certaines langues. Il

y a deux moyens empiriques pour estimer la probabilite de la causalite externe: a) en

observant la distribution typologique des morphemes causatifs vs. anticausatifs dans la

structure des verbes causatifs lexicaux au travers d’un grand nombre des langues et b)

en observant la distribution de frequences des realisations transitives vs. intransitives

des verbes dans un corpus d’une langue individuelle. Notre etude montre que ces deux

mesures sont correlees. En appliquant la mesure basee sur le corpus, la position sur

l’echelle de la causalite externe peut etre determinee automatiquement pour un grand

nombre de verbes.

Le sujet de la troisieme etude de cas est la relation entre les deux caracteristiques tem-

porales des evenements encodees par la categorie grammaticale d’aspect verbale: la

longueur et la delimitation temporelle. L’etude montre que ces deux caracteristiques

interagissent d’une maniere complexe mais previsible, ce qui est a l’origine de la vari-

ation observee dans les realisations morphosyntactiques des verbes. Les intuitions des

locuteurs natifs anglais sur la longueur possible d’un evenement decrit par un verbe

(court vs. long) peuvent etre predites sur la base du marquage formel d’aspect verbal

dans les verbes correspondants serbes. L’exactitude des predictions basees sur le modele

bi-linguistique est superieure a la performance du meilleur modele monolanguistique.

Une parmi les contributions principales de cette these est la nouvelle methodologie

experimentale qui se base sur le traitement automatique des corpus paralleles et sur

l’inference statistique. Les trois caracteristiques semantiques des evenements decrits

par des verbes (la dynamique, la probabilite de la causalite externe, la longueur) sont

inferees empiriquement a partir d’observations extraites automatiquement des grands

corpus paralleles (contenant jusqu’a plus d’un million de phrases pour chaque langue)

automatiquement analyses et alignes. Les generalisations generalisations sont acquises

de donnees de corpus de maniere automatique en utilisant l’inference statistique et les

techniques d’apprentissage automatique. L’exactitude des predictions effectuees sur la

base des generalisations est estimee de maniere experimentale en utilisant un echantillon

separe de donnees de test.

7

Acknowledgements

This dissertation has greatly benefited from the help and support of numerous friends

and colleagues and I wish to express my gratitude to all of them here.

First and foremost, I would like to thank my supervisor, Paola Merlo, for the com-

mitment with which she has supervised this dissertation, for sharing generously her

knowledge and experience in countless hours spent discussing my work and reading my

pages, for treating my ideas with care and attention, and for showing me that I can do

better than I thought I could.

I am most thankful to Vesna Polovina and Jacques Mœschler, who made it possible for

me to move from Belgrade to Geneva and who have discretely looked after me throughout

my studies.

I thank Balthasar Bickel, Jonas Kuhn, and Martha Palmer, who kindly agreed to be

members of the defence committee, and to Jacques Mœschler, who agreed to be the

president of the jury.

I have gathered much of the knowledge and skills necessary for carrying out this research

in the discussions and joint work with Boban Arsenijevic, Effi Georgala, Andrea Ges-

mundo, Kristina Gulordava, Maja Milicevic, Lonneke van der Plas, Marko Simonovic,

and Balsa Stipcevic. I am thankful for the time they spent working and thinking with

me.

I appreciate very much the assistance of James Henderson, Jonas Kuhn, and Gerlof

Bouma, who shared their data with me, allowing me to spend less time processing

corpora, so I could spend more time thinking about the experiments.

9

I am thankful to my colleagues in the Department of General Linguistics in Belgrade,

in the Linguistics Department in Geneva, and in the CLCL research group for their

kindness and support. On various occasions, I felt lucky to be able to talk to Tijana

Asic, Lena Baunaz, Anamaria Bentea, Frederique Berthelot, Giuliano Bocci, Eva Cap-

itao, Maja Djukanovic, Nikhil Garg, Jean-Philippe Goldman, Asheesh Gulati, Tabea

Ihsane, Borko Kovacevic, Joel Lang, Antonio Leoni de Leon, Gabriele Musillo, Goljihan

Kashaeva, Alexis Kauffmann, Christopher Laenzlinger, Jasmina Moskovljevic Popovic,

Luka Nerima, Natalija Panic Cerovski, Genoveva Puskas, Lorenza Russo, Yves Scherrer,

Violeta Seretan, Gabi Soare, Zivka Stojiljkovic, Eric Wehrli, and Richard Zimmermann.

I would also like to thank Pernilla Danielsson, who helped me start doing computa-

tional linguistics while I was a visiting student at the Centre for Corpus Research at the

University of Birmingham.

In the end, I would like to express my gratitude to Fabio, who has stayed by my side

despite all the evenings, weekends, and holidays dedicated to this dissertation.

10

Contents

1. Introduction 21

1.1. Grammatically relevant components of the meaning of verbs . . . . . . . 22

1.2. Natural language processing in linguistic research . . . . . . . . . . . . . 24

1.3. Using parallel corpora to study language variation . . . . . . . . . . . . . 25

1.4. The overview of the dissertation . . . . . . . . . . . . . . . . . . . . . . . 29

2. Overview of the literature 33

2.1. Theoretical approaches to the argument structure . . . . . . . . . . . . . 34

2.1.1. The relational meaning of verbs . . . . . . . . . . . . . . . . . . . 36

2.1.2. Atomic approach to the predicate-argument structure . . . . . . . 38

2.1.3. Decomposing semantic roles into clusters of features . . . . . . . . 42

Proto-roles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

The Theta System . . . . . . . . . . . . . . . . . . . . . . . . . . 45

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

2.1.4. Decomposing the meaning of verbs into multiple predicates . . . . 49

Aspectual event analysis . . . . . . . . . . . . . . . . . . . . . . . 50

Causal event analysis . . . . . . . . . . . . . . . . . . . . . . . . . 52

2.1.5. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

2.2. Verb classes and specialised lexicons . . . . . . . . . . . . . . . . . . . . . 54

2.2.1. Syntactic approach to verb classification . . . . . . . . . . . . . . 54

2.2.2. Manually annotated lexical resources . . . . . . . . . . . . . . . . 57

FrameNet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

The Proposition Bank (PropBank) . . . . . . . . . . . . . . . . . 62

VerbNet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

Comparing the resources . . . . . . . . . . . . . . . . . . . . . . . 67

11

Contents

2.3. Automatic approaches to the predicate-argument structure . . . . . . . . 70

2.3.1. Early analyses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

2.3.2. Semantic role labelling . . . . . . . . . . . . . . . . . . . . . . . . 73

Standard semantic role labelling . . . . . . . . . . . . . . . . . . . 73

Joint and unsupervised learning . . . . . . . . . . . . . . . . . . . 80

2.3.3. Automatic verb classification . . . . . . . . . . . . . . . . . . . . . 81

2.4. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

3. Using parallel corpora for linguistic research — rationale and methodology 87

3.1. Cross-linguistic variation and parallel corpora . . . . . . . . . . . . . . . 88

3.1.1. Instance-level microvariation . . . . . . . . . . . . . . . . . . . . . 89

3.1.2. Translators’ choice vs. structural variation . . . . . . . . . . . . . 92

3.2. Parallel corpora in natural language processing . . . . . . . . . . . . . . . 94

3.2.1. Automatic word alignment . . . . . . . . . . . . . . . . . . . . . . 94

3.2.2. Using automatic word alignment in natural language processing . 98

3.3. Statistical analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

3.3.1. Summary tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

3.3.2. Statistical inference and modelling . . . . . . . . . . . . . . . . . 103

3.3.3. Bayesian modelling . . . . . . . . . . . . . . . . . . . . . . . . . . 108

3.4. Machine learning techniques . . . . . . . . . . . . . . . . . . . . . . . . . 112

3.4.1. Supervised learning . . . . . . . . . . . . . . . . . . . . . . . . . . 113

3.4.2. Unsupervised learning . . . . . . . . . . . . . . . . . . . . . . . . 120

3.4.3. Learning with Bayesian Networks . . . . . . . . . . . . . . . . . . 125

3.4.4. Evaluation of predictions . . . . . . . . . . . . . . . . . . . . . . . 127

3.5. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

4. Force dynamics schemata and cross-linguistic alignment of light verb con-

structions 131

4.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

4.2. Theoretical background . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

4.2.1. Light verb constructions as complex predicates . . . . . . . . . . . 134

4.2.2. The diversity of light verb constructions . . . . . . . . . . . . . . 138

12

Contents

4.3. Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

4.3.1. Experiment 1: manual alignment of light verb constructions in a

parallel corpus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144

Materials and methods . . . . . . . . . . . . . . . . . . . . . . . . 145

Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . 148

4.3.2. Experiment 2: Automatic alignment of light verb constructions in

a parallel corpus . . . . . . . . . . . . . . . . . . . . . . . . . . . 150



4.4. General discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

4.4.1. Two force dynamics schemata in light verbs . . . . . . . . . . . . 159

4.4.2. Relevance of the findings to natural language processing . . . . . 161

4.5. Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162

4.6. Summary of contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 165

5. Likelihood of external causation and the cross-linguistic variation in lexical

causatives 167

5.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

5.2. Theoretical accounts of lexical causatives . . . . . . . . . . . . . . . . . . 171

5.2.1. Externally and internally caused events . . . . . . . . . . . . . . . 172

5.2.2. Two or three classes of verb roots? . . . . . . . . . . . . . . . . . 174

5.2.3. The scale of spontaneous occurrence . . . . . . . . . . . . . . . . 176

5.3. Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179

5.3.1. Experiment 1: Corpus-based validation of the scale of spontaneous

occurrence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181



5.3.2. Experiment 2: Scaling up . . . . . . . . . . . . . . . . . . . . . . 186



5.3.3. Experiment 3: Spontaneity and cross-linguistic variation . . . . . 190



13

Contents

5.3.4. Experiment 4: Learning spontaneity with a probabilistic model . 202

The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204

Experimental evaluation . . . . . . . . . . . . . . . . . . . . . . . 206


5.4.1. The scale of external causation and the classes of verbs . . . . . . 212

5.4.2. Cross-linguistic variation in English and German . . . . . . . . . 213


5.5. Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215


6. Unlexicalised learning of event duration using parallel corpora 221

6.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221

6.2. Theoretical background . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226

6.2.1. Aspectual classes of verbs . . . . . . . . . . . . . . . . . . . . . . 226

6.2.2. Observable traits of verb aspect . . . . . . . . . . . . . . . . . . . 231

6.2.3. Aspect encoding in the morphology of Serbian verbs . . . . . . . . 233

6.3. A quantitative representation of aspect based on cross-linguistic data . . 238

6.3.1. Corpus and processing . . . . . . . . . . . . . . . . . . . . . . . . 240

6.3.2. Manual aspect classification in Serbian . . . . . . . . . . . . . . . 243

6.3.3. Morphological attributes . . . . . . . . . . . . . . . . . . . . . . . 244

6.3.4. Numerical values of aspect attributes . . . . . . . . . . . . . . . . 245

6.4. Experiment: Learning event duration with a statistical model . . . . . . 248

6.4.1. The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249

The Bayesian net classifier . . . . . . . . . . . . . . . . . . . . . . 251

6.4.2. Experimental evaluation . . . . . . . . . . . . . . . . . . . . . . . 253




6.5.1. Aspectual classes . . . . . . . . . . . . . . . . . . . . . . . . . . . 260


6.6. Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262


14

Contents

7. Conclusion 265

7.1. Theoretical contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . 267

7.2. Methodological contribution . . . . . . . . . . . . . . . . . . . . . . . . . 269

7.3. Directions for future work . . . . . . . . . . . . . . . . . . . . . . . . . . 270

Bibliography 273

A. Light verb constructions data 293

A.1. Word alignment of the constructions with ’take’ . . . . . . . . . . . . . . 293

A.2. Word alignment of the constructions with ’make’ . . . . . . . . . . . . . 297

A.3. Word alignments of regular constructions . . . . . . . . . . . . . . . . . . 301

B. Corpus counts and measures for lexical causatives 305

C. Verb aspect and event duration data 317

15

List of Figures

1.1. Cross-linguistic mapping between morphosyntactic categories. . . . . . . 26

1.2. Cross-linguistic mapping between morphosyntactic categories. . . . . . . 28

3.1. Word alignment in a parallel corpus . . . . . . . . . . . . . . . . . . . . . 95

3.2. Probability distributions of the morphological forms and syntactic reali-

sations of the example instances. . . . . . . . . . . . . . . . . . . . . . . 105

3.3. Probability distributions of the example verbs and their frequency. . . . . 106

3.4. A general graphical representation of the normal distribution. . . . . . . 107

3.5. An example of a decision tree . . . . . . . . . . . . . . . . . . . . . . . . 116

3.6. An example of a Bayesian network . . . . . . . . . . . . . . . . . . . . . 126

4.1. A schematic representation of the structure of a light verb construction

compared with a typical verb phrase . . . . . . . . . . . . . . . . . . . . 132

4.2. Constructions with vague action verbs . . . . . . . . . . . . . . . . . . . 143

4.3. True light verb constructions . . . . . . . . . . . . . . . . . . . . . . . . . 144

4.4. Extracting verb-noun combinations . . . . . . . . . . . . . . . . . . . . . 146

4.5. The difference in automatic alignment depending on the direction. . . . . 152

4.6. The distribution of nominal complements in constructions with take . . . 155

4.7. The distribution of nominal complements in constructions with make . . 155

4.8. The distribution of nominal complements in regular constructions . . . . 156

4.9. The difference in automatic alignment depending on the complement fre-

quency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

5.1. The correlation between the rankings of verbs on the scale of spontaneous

occurrence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185

5.2. Density distribution of the Sp value in the two samples of verbs . . . . . 188

17

List of Figures

5.3. Collecting data on lexical causatives . . . . . . . . . . . . . . . . . . . . . 193

5.4. Density distribution of the Sp value over instances of 354 verbs . . . . . . 198

5.5. Joint distribution of verb instances in the parallel corpus . . . . . . . . . 201

5.6. Bayesian net model for learning spontaneity. . . . . . . . . . . . . . . . . 204

5.7. The Interaction of the factors involved in the causative alternation . . . . 211

6.1. Traditional lexical verb aspect classes, known as Vendler’s classes . . . . 227

6.2. Serbian verb structure summary . . . . . . . . . . . . . . . . . . . . . . . 237

6.3. Bayesian net model for learning event duration . . . . . . . . . . . . . . . 251

18

List of Tables

2.1. Frame elements for the verb achieve . . . . . . . . . . . . . . . . . . . . . 61

2.2. Some combinations of frame elements for the verb achieve. . . . . . . . . 62

2.3. The PropBank lexicon entry for the verb pay. . . . . . . . . . . . . . . . 65

2.4. The VerbNet entry for the class Approve-77. . . . . . . . . . . . . . . . . 66

3.1. Examples of instance variables . . . . . . . . . . . . . . . . . . . . . . . . 102

3.2. Examples of type variables . . . . . . . . . . . . . . . . . . . . . . . . . . 102

3.3. A simple contingency table summarising the instance variables . . . . . . 102

3.4. An example of data summary in Bayesian modelling . . . . . . . . . . . . 109

3.5. An example of a data record suitable for supervised machine learning . . 114

3.6. Grouping values for training a decision tree . . . . . . . . . . . . . . . . . 118

3.7. An example of a data record suitable for supervised machine learning . . 120

3.8. An example of probability estimation using the expectation-maximisation

algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

3.9. Precision and recall matrix . . . . . . . . . . . . . . . . . . . . . . . . . . 127

4.1. Types of mapping between English constructions and their translation

equivalents in German. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

4.2. Well-aligned instances of light verb constructions . . . . . . . . . . . . . 153

4.3. The three types of constructions partitioned by the frequency of the com-

plements in the sample. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154

4.4. Counts and percentages of well-aligned instances in relation with the fre-

quency of the complements in the sample . . . . . . . . . . . . . . . . . . 158

5.1. Cross-linguistic variation in lexical causatives . . . . . . . . . . . . . . . 169

5.2. Morphological marking of cause-unspecified verbs . . . . . . . . . . . . . 175

19

List of Tables

5.3. Morphological marking across languages . . . . . . . . . . . . . . . . . . 177

5.4. An example of an extracted instance of an English alternating verb and

its translation to German . . . . . . . . . . . . . . . . . . . . . . . . . . 195

5.5. Examples of parallel instances of lexical causatives. . . . . . . . . . . . . 196

5.6. Contingency tables for the English and German forms in different samples

of parallel instances. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200

5.7. Examples of the cross-linguistic input data . . . . . . . . . . . . . . . . . 207

5.8. Agreement between corpus-based and typology-based classification of verbs.

The classes are denoted in the following way: a=anticausative (interanally

caused), c=causative (externally caused) , m=cause-unspecified. . . . . . 208

5.9. Confusion matrix for monolingual and cross-linguistic classification on 2

classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210

5.10. Confusion matrix for monolingual and cross-linguistic classification on 3

classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210

6.1. A relationship between English verb tenses and aspectual classes. . . . . 231

6.2. Serbian lexical derivations . . . . . . . . . . . . . . . . . . . . . . . . . . 234

6.3. Serbian lexical derivations with a bare perfective . . . . . . . . . . . . . . 237

6.4. An illustration of the MULTEX-East corpus . . . . . . . . . . . . . . . . 241

6.5. A sample of the verb aspect data set. . . . . . . . . . . . . . . . . . . . . 248

6.6. A sample of the two versions of data . . . . . . . . . . . . . . . . . . . . 256

6.7. Results of machine learning experiments . . . . . . . . . . . . . . . . . . 258

20

1. Introduction

Languages use different means to express the same content. Variation in the choice of

lexical items or syntactic constructions is possible without changing the meaning of a

sentence. For example, any of the sentences in (1.1a-c) can be used to express the same

event. Similarly, the meaning of the sentences in (1.2a-b), (1.3a-b), and (1.4a-b) can

be considered as equivalent. The sentences in (1.1) illustrate the variation in the choice

of lexical items, while the sentences in (1.2-1.4) show that the syntactic structure of a

sentence can be changed without changing the meaning. In both cases, the variation

is limited to the options which are provided by the rules of grammar. In order to be

exchangeable, linguistic units have to share certain properties. Identifying the properties

shared by different formal expressions of semantically equivalent units is, thus, a way of

identifying abstract elements of the structure of language.

(1.1) a. Mary drank a cup of tea.

b. Mary took a cup of tea.

c. Mary had a cup of tea.

d. Mary had a cup of coffee.

As illustrated in (1.1d), verbs allow alternative expressions more easily than other cat-

egories. Replacing the noun tea, for example, by coffee changes the meaning of the

sentence so that (1.1d) can no longer be considered as equivalent with (1.1a-c). The

property which allows verbs to alternate more easily than other categories is their re-

lational meaning. In the given examples, the verbs drink, take, and have relate the

nouns Mary and tea. The relational meaning of a verb is commonly represented as the

predicate-argument structure, where a verb is considered as a predicate which takes other

21

1. Introduction

constituents of a sentence as its arguments. The number and the type of the arguments

that a verb takes in a particular instance is partially determined by the verb’s meaning

and partially by the contextual and pragmatic factors involved in the instance.

(1.2) a. Mary laughed.

b. Mary had a laugh.

(1.3) a. Adam broke the laptop.

b. The laptop broke.

(1.4) a. John pushed the cart.

b. John pushed the cart for some time.

In this dissertation, we study systematic variation in the use of verbs involving alterna-

tion in the syntactic structure, as in (1.2-1.4). We study frequency distributions of the

syntactic alternants as an observable indicator of the underlying meaning of verbs with

the aim of discovering the components of verbs’ meaning which are relevant for their

predicate-argument structure and for the grammar of language.

1.1. Grammatically relevant components of the meaning

of verbs

As argued by Pesetsky (1995) and later by Levin and Rappaport Hovav (2005), only some

of the potential components of the meaning of verbs are grammatically relevant. For

example, the distinction between verbs describing loud speaking (e.g. shout) and verbs

describing quiet speaking (e.g. whisper) is grammatically irrelevant in the sense that it

does not influence any particular syntactic behaviour of these verbs (Pesetsky 1995).

Contrary to this, the distinction between verbs which describe primarily the manner of

speaking (whisper) and verbs which describe primarily the content of speaking (e.g. say)

is grammatically relevant in the sense that the latter group of verbs can be used without

the complementizer that, while the former cannot. Along the same lines, Levin and

22

1.1. Grammatically relevant components of the meaning of verbs

Rappaport Hovav (2005) argue that the quality of sound described by verbs of sound

emission — volume, pitch, resonance, duration — does not influence their syntactic

behaviour. Syntactic behaviour of these verbs is, in fact, influenced by the source of the

sound: verbs which describe sound emission with the source of the sound external to the

emitting object (e.g. rattle) can alternate between transitive and intransitive uses (in a

similar fashion as break in (1.3)), while verbs which describe sound emission with the

source of the sound internal to the emitting object (e.g. rumble) do not alternate.

Our research continues in the same direction investigating other semantic properties of

verbs which are potentially relevant for the grammar. We take into consideration a

wide range of verbs and their syntactic realisations. If a particular observed distribution

of syntactic alternants can be predicted from a semantic poperty of a verb, then we

can say that this property underlies the distribution. If a semantic property undrlies a

frequency distribution of syntactic alternants, then this property can be considered as

grammaticaly relevant.

We focus on three kinds of alternations in realisation of verbs’ arguments. First, by

studying the alternation between light verb constructions (1.2b) and the corresponding

single verbs (1.2a), we address the issue of whether certain lexical content, in the form of

the predicate-argument structure, is present in the verbs which are used as light verbs,

such as have in (1.2b). Determining whether some components of meaning are present

in light verbs is important for understanding whether the choice of the light verb in

a construction is arbitrary or it is constrained by the meaning of light verbs. Second,

we study the alternation in the use of lexical causatives such as break in (1.3). Lexical

causatives are the verbs which can be used in two ways: as causative (1.3a), where the

agent or the causer of the event described by the verb is realised as a constituent of a

sentence, and as anticausative (1.3b), where the agent or the causer is not syntactically

realised. Many verbs across many different languages can alternate in this way. However,

the fact that some verbs in some languages do not alternate raises the question which

is addressed in this dissertation: What property of verbs is responsible for allowing or

blocking the alternation? Finally, we study the factors involved in the interpretation of

temporal properties of events described by verbs. As illustrated in (1.4), the temporal

properties of events described by verbs play a role in syntactic structuring of a sentence.

For example, the event of pushing is interpreted as short by default (1.4a). With the

23

1. Introduction

appropriate temporal modifier, as in (1.4b), it can also be interpreted as lasting for a

longer time. In contrast to this, other verbs, such as tick, stay, walk, describe events

which are understood as lasting for some time by default. We look for observable indica-

tors in the use of a wide range of verbs pointing to the event duration which is implicit

to their meaning.

1.2. Natural language processing in linguistic research

The approach that we take in addressing the defined questions is empirical and com-

putational. We take advantage of automatic language processing to collect and analyse

large data sets, applying established statistical approaches to infer elements of linguistic

structure from the patterns in the observed variation. The tools, methods, and re-

sources which we use are originally developed for practical natural language processing

tasks which fall within the domain of computational linguistics. The developments in au-

tomatic language processing are directly related to the increasing demand for automatic

analysis of large amounts of linguistic contents which are now freely available (mostly

through the Internet). Natural language processing tasks include automatic informa-

tion extraction, question answering, translation etc. Despite the fact that it provides

extremely rich resources for empirical linguistic investigations, natural language pro-

cessing technology has rarely been used for theoretical linguistic research. On the other

hand, linguistic representations that are used in developing language technology rarely

reflect the current state-of-the-art in linguistic theory. Our research should contribute

to bridging the gap between theoretical and computational linguistics by addressing

current theoretical discussion with a computational methodology.

The work in this dissertation draws on the work in natural language processing in two

ways. First, we use automatic processing tools to extract the information from large

language corpora. For example, to identify syntactic forms of the realisations of verbs, we

use automatically parsed corpora. The information provided by the parses is then used to

extract automatically the instances which are relevant for a particular question. Second,

we use natural language processing methodology to analyse the extracted instances. This

methodology involves three main components: a) the generalisations in the observations

24

1.3. Using parallel corpora to study language variation

are captured by designing statistical models; b) the parameters of the models are learnt

automatically from the extracted data applying machine learning techniques; c) the

predictions of the models are tested on an independent set of data, quantifying and

measuring the performance. Adopting this methodology for our research allows us not

only to study language use in a valid experimental framework, but also to discover

generalisations which can be integrated into further development of natural language

processing more easily than the generalisations based on linguistic introspection.


Our approach to the relationship between the variation in language use and the structure

of language takes into account both language-internal and cross-linguistic variation. This

is achieved by extracting verb instances from parallel corpora. By studying the variation

in the use of verbs in parallel corpora, we combine and extend two main approaches to

language variation: the corpus-based approach to language-internal variation and the

theoretical approach to cross-linguistic variation.

Corpus-based studies of linguistic variation have been mostly monolingual, following the

use of linguistic units either over a period of time or across different language registers.

Extending the corpus-based approach to parallel corpora allows a better insight into

structural linguistic elements, setting them apart from other potential factors of varia-

tion. Consider, for example, the alternations in (1.2-1.4). An occurrence of one or the

other syntactic alternant in a monolingual corpus depends partially on the predicate-

argument structure of the verbs and partially on the contextual and pragmatic factors.

However, if we can observe actual translations of the sentences, then we can observe at

least two uses of semantically equivalent units in the same contextual and pragmatic

conditions, since these conditions are constant in translation. In this way, we control

for contextual and pragmatic factors while potentially observing the variation due to

structural factors.

Unlike language-internal variation, which has become the subject of research relatively

recently, with the development of corpus-based approaches, cross-linguistic variation is

25

1. Introduction

traditionally one of the core issues in theoretical linguistics. Differences in the expres-

sions of the same contents across languages have always been analysed with the aim of

discovering universally invariable elements of the structure of language which constrain

the variation. Consider, for example, the English sentence in (1.5a) and its corresponding

German, Serbian, and French sentences in (1.5b-d).

(1.5) a. Mary has just sent the letter. (English)

b. Maria hat gerade eben den Brief geschickt. (German)

c. Marija je upravo poslala pismo. (Serbian)

d. Marie vein d’envoyer la lettre. (French)

English German Serbian French

presentperfect

adverb+perfect

prefixvenir+infinitive

Figure 1.1.: Cross-linguistic mapping between morphosyntactic categories.

All the four sentences describe a short completed action that happened immediately

before the time in which the sentence is uttered, but the meaning of shortness, com-

pleteness, and time (immediate precedence) is expressed in different ways in the four

languages. In English, this meaning is encoded with a verb tense, present perfect.

German uses more general perfect tense, and the immediate precedence component is

encoded in the adverbs (gerade eben). French, on the other hand, does not use any

particular verb conjugation to express this meaning, but rather a construction which

consists of a semantically impoverished verb (venir ’come’) and the main verb (envoyer

’send’) in the neutral, infinitive form. The corresponding Serbian expression is formed

in yet another different way: through lexical derivation. The verb poslati used in (1.5c)

is derived from the verb slati, which does not encode any specific temporal properties,

by adding the prefix po-. Figure 1.1 summarises the identified grammatical mappings

across languages. Note also that, unlike the sentences in other languages, the French

sentence does not contain a temporal adverb. The meaning of immediate precedence

26


is already encoded as part of the meaning of the constructions formed with the verb

’venir’.

These examples illustrate systematic variation across languages, and not just incidental

differences between these particular sentences. If we replace the constituents of the

sentences with some other members of their paradigms, we will observe the same patterns

of variation. For instance, we can replace the phrase send the letter in the English

sentence and its lexical counterparts in German, Serbian, and French by some other

phrases, such as open the window, read the message, arrive to the meeting and so on.

The choice of the corresponding morphosyntactic categories can be expected to stay

the same. The regular patterns in cross-linguistic variation are due to the fact that

sentences are composed of the same abstract units. As mentioned before, all the four

sentences in (1.5) express the same event, with the same temporal properties (shortness,

completeness, immediate precedence). The fact that they influence (morpho)syntactic

realisations of verbs makes these properties grammatically relevant. The fact that they

are equally interpreted across languages, despite the differences in the morphosyntactic

realisations, makes them candidates for universal elements of the structure of language.

Theoretical approaches to cross-linguistic variation are concerned with identifying not

only the elements of linguistic structure which are invariable across languages, but also

the parameters of variation and their possible settings. With these two elements one

could then construct a general representation of language capacity shared by all speakers

of all languages. In this system, the grammar of any particular language instantiates the

general grammar by setting the parameters to a certain value. For example, temporal

properties of events in our example, which are invariable across languages, can be en-

coded in a syntactic construction (French), in the morphology (English), or in the lexical

derivation (Serbian). Ideally, the number of possible values for a parameter should be

small.

However, identifying the parameters of cross-linguistic variation and their possible set-

tings is far from being a trivial task. Even though there are some regular patterns of

cross-linguistic mapping, as we saw earlier, it is hard to define general rules which apply

to all instances of a given category, independently of a given context. In fact, when we

take a closer look, finding regularities in cross-linguistic variation turns out to be a very

27

1. Introduction

difficult task for which no common methodology has been proposed. To illustrate the

difficulties, we will look again at the example of English present perfect tense, for which

we have defined cross-linguistic mappings shown in Figure 1.1. As we can see in (1.6),

the mappings in Figure 1.1 do not hold for all the instances of English present perfect

tense. A different use of this tense in English brings to rather different mappings.

(1.6) a. Mary still has not seen the film. (English)

b. Maria hat noch immer nicth den Film gesehen. (German)

c. Marija jos nije gledala film. (Serbian)

d. Marie n’a pas encore vu le film. (French)

English German Serbian French

presentperfect

perfectbareform

passecompose

Figure 1.2.: Cross-linguistic mapping between morphosyntactic categories.

Figure 1.2 summarises the mappings between the sentences in (1.6). We can see that,

iinstead of the construction with the verb venir, the corresponding French form n this

case is a verb tense (passe compose). The corresponding Serbian verb in this context is

neither prefixed nor perfective. This means that the English present perfect tense has

multiple cross-linguistic mappings even in this small sample of only two other languages

(the German form can be considered invariable in this case). Other uses might be

mapped in yet different ways. For instance, there can be a use which maps to French

as in Figure 1.2, and to Serbian as in Figure 1.1. If we take into account all the other

languages and all possible uses of present perfect tense in English, the number of possible

cross-linguistic mappings of this single morphological category is likely to become very

big. We can expect to encounter the same situation with all the other categories and

their combinations. This creates a very large space of possible cross-linguistic mappings,

which is hard to explore and to account for in an exhaustive fashion.

28

1.4. The overview of the dissertation

Extracting verb instances from parallel corpora allows us to observe directly a wide range

of cross-linguistic mappings of the target morphosyntactic categories at the instance

level, taking into account contextual factors. With a large number of instances analysed

using computational and statistical methods, we can take a new perspective on the

cross-linguistic variation. Zooming out to analyse general tendencies in the data, rather

than individual cases, we can identify patterns signalling potential constraints on the

variation. Even though this approach is not exhaustive, it is systematic in the sense

that it allows us to observe patterns in cross-linguistic variation in large samples and

to use statistical inference to formulate generalisations which hold beyond the observed

samples.


The dissertation consists of seven chapters. In addition to Introduction and Conclusion,

there are five central chapters which are divided between two main parts. The first

part (Chapters 2 and 3) presents the conceptual and technical background of our work,

the rationale for our methodological choices, as well as a detailed description of general

methods used in our experiments. The second part (Chapters 4, 5, and 6) contains three

case studies in which our experimental methodology is used to address three specific

theoretical questions.

In Chapter 2, we discuss the issues in the predicate-argument structure of verbs from

two points of view: theoretical and computational. The theoretical track follows the

development in the view of the predicate argument structure from the first proposals

which divide the grammatical and the idiosyncratic components of the lexical structure

of verbs to the current view of verbs as composed of multiple predicates, which is adopted

in our research. We review theoretical arguments for abandoning the initial “atomic”

view of the predicate-argument structure, as well as some proposals for its systematic

decomposition into smaller components. We then proceed by reviewing the work on

extensive verb classification, which relates the grammatical and the idiosyncratic layer of

the lexical structure of verbs. We discuss the principles of semantic classification of verbs

on the basis of their syntactic behaviour, as well as practical implementations of verb

29

1. Introduction

classification principles in developing extensive language resources. Finally, we review

approaches to automatic acquisition of verb classification and the predicate-argument

structure, discussing the representations and methods used for these tasks.

Chapter 3 deals with the methodology of using parallel corpora for linguistic research.

Since parallel corpora are not commonly used as a source of data for linguistic research,

we first present our rationale for this choice, discussing its advantages, but also its

limitations. We then give an overview of natural language processing approaches based

on parallel corpora and the contributions of this line of research. The second part of the

chapter deals with the technical and practical issues in using natural language processing

methodology for linguistic research. We first describe steps in processing parallel corpora

for extracting linguistic data, in particular, automatic word alignment, which is crucial

for our approach. We then turn to the methods used for analysing the extracted data

providing the technical background necessary to follow the discussion in the three case

studies. The background includes an introduction to statistical inference and modelling

in general, as well as to Bayesian modelling in particular, which is followed by an overview

of four standard machine learning classification techniques which are used or referred

to in our case studies: naıve Bayes, decision tree, Bayesian net, and the expectation-

maximisation algorithm.

The first case study, on light verb constructions, is presented in Chapter 4. We first

give an overview of the theoretical background and the questions raised by light verb

constructions. We introduce two classes of light verb constructions discussed in the liter-

ature, true light verb constructions and constructions with vague action verbs. We then

introduce our proposed classification which is based on verb types. We argue that light

verb constructions headed by light take behave like true light verb constructions, while

the constructions headed by light make behave like the constructions with vague action

verbs. We relate this behaviour to the force dynamics representation of the predicate-

argument structure of these verbs. We then present two experiments in which we test

two hypotheses about the relationship between the force dynamics in the meaning of the

verbs and the cross-linguistic frequency distribution of the alternating morhosyntactic

forms.

The case study on the causative alternation is presented in Chapter 5. We start by

30


reviewing the proposed generalisations addressing the meaning of the verbs which par-

ticipate in the causative alternation. In particular, we address the notions of change of

state, external vs. internal causation, and cross-linguistic variation in the availability of

the alternation. We then introduce the discussion on the number of classes into which

verbs should be classified with respect to these notions. Two proposal have been put

forward in the literature: a) a two-way distinction between alternating and not alter-

nating verbs, where alternating verbs are characterised as describing externally caused

events, while the verbs which do not alternate describe internally caused events; b) a

three-way classification involving a third class of verbs situated between the two previ-

ously proposed classes. We then discuss the distribution of the morphological marking

on alternating verbs across languages as a potential indicator of the grammatically rel-

evant meaning of the alternating verbs. This leads us to introducing the notion of the

likelihood of external causation. The experimental part of this study consists of four

steps. In the first step, we validate a corpus based measure of the likelihood of external

causation showing that it correlates with the typological distribution of the morphologi-

cal marking. In the second step, we show that the corpus based measure can be extended

to a large sample of verbs. In the third step, we extract the instances of the large sam-

ple of verbs from a parallel corpus and test the influence of the likelihood of external

causation on the cross-linguistic distribution of their morphosyntactic realisations. In

the fourth step, we address the issue of classifying the alternating verbs by designing a

statistical model which takes as input cross-linguistic realisations of verbs and outputs

their semantic classification. We test the model in two modes, on the two-way and on

the three-way classification.

The last case study, presented in Chapter 6, deals with the representation of grammati-

cally relevant temporal properties of events described by a wide range of verbs. We start

by introducing verb aspect as a grammatical category usually thought to encode tempo-

ral meaning. More specifically, we discuss two notions related to verb aspect: temporal

boundedness and event duration. We then discuss Serbian verb derivations associated

with verb aspect as a potential observable indicator of these two temporal properties

of events described by verbs. We proceed by proposing a quantitative representation of

Serbian verb aspect based on cross-linguistic realisations of verbs extracted from par-

allel corpora. We then design a Bayesian model which predicts the duration (short vs.

31

1. Introduction

long) of events taking this representation as input. We test the performance of the

model against English native speakers’ judgments of the duration of events described by

English verbs. We compare our results to the results of models based on monolingual

English input.

In Chapter 7, we draw some general conclusions, pointing to the limitations of the

current approach as well as to some directions for future research.

32

2. Overview of the literature

The conceptual and methodological framework of the experiments presented in this dis-

sertation encompasses three partially interrelated lines of research: theoretical accounts

of the grammatically relevant meaning of verbs, its extensive descriptions in specialised

lexicons, and its automatic acquisition from language corpora.

Theoretical accounts of the meaning of verbs are crucial for defining the hypotheses

which are tested in our experiments. Our hypotheses are formulated in the context and

framework of recent developments in theoretical accounts of lexical representation of

verbs. While using the tools and the methodology developed in computational linguis-

tics, our main goal is not to develop a new tool or resource, but to extend the general

knowledge about what kinds of meaning are actually part of the lexical representation

of verbs and how they are related to the grammar. Our work is related to the work

on constructing comprehensive specialised lexicons of verbs because we work with large

sets of verbs assigning specific lexical and grammatical properties to each verb in each

sample. Finally, we follow the work on automatic acquisition of the meaning of verbs

in that we learn the elements of their lexical representation automatically from the ob-

served distributions of their realisations in a corpus. This aspect distinguishes our work

from theoretical approaches, as well as from the work on developing specialised lexicons,

which are based on linguistic introspection rather than on empirical observations.

This chapter contains an overview of the existing research in all three domains. In Sec-

tion 2.1, we follow the developments in theoretical approaches to the meaning of verbs.

We start by introducing the notion of predicate-argument structure of verbs, discussing

its role in the grammar of language, as well as in linguistic theory (2.1.1). We proceed by

reviewing proposed theoretical accounts which represent general views of the predicate-

argument structure in the literature, discussing at length crucial turning points in the

33


theoretical development leading to the temporal and causal decomposition of the mean-

ing of verbs which is adopted in our experiments. In Section 2.2, we discuss the principles

of large-scale implementations of some views of the predicate-argument structure. We

summarise the main ideas behind the syntactic behavioural approach to the meaning

of verbs (2.2.1), which is followed by descriptions of three lexical resources which con-

tain thousands of verbs with explicit analyses of their predicate-argument structure. In

Section 2.3, we discuss approaches to automatic acquisition of the predicate-argument

structure from language corpora which rely on the described lexical resources conceptu-

ally (they adopt the principles of syntactic approach to verb meaning) and practically

(they use the resources for training and testing systems for automatic acquisition).

2.1. Theoretical approaches to the argument structure

It is generally assumed in linguistic theory that the structure of a sentence depends,

to a certain degree, on the meaning of its main verb. Some verbs, such as see in

(2.1) require a subject and an object; others, such as laugh in (2.2), form grammatical

sentences expressing only the subject; others, such as tell in (2.3) require expressing

three constituents. (Clauses with more than three principal constituents are rare.) The

assumption concerning these observations is that the association of certain verbs with a

certain number and kind of constituents is not due to chance, but that it is part of the

grammar of language.

(2.1) [Mary]subject

saw [a friend].object

(2.2) [Mary]subject

laughed.

(2.3) [Mary]subject

told [her friend]indirect-object

[a story].object

34


Although the relation between the meaning of verbs and the available syntactic patterns

seems obvious, defining precise rules to derive a phrase structure from the lexical struc-

ture of a verb proves to be a difficult task. The task, known in the the linguistic literature

as the linking problem, is one of the central concerns of the theory of language (Baker

1997). The main difficulty in linking the meaning of verbs and the form of the phrases

that they head is in analysing verbs’ meaning so that the components responsible for

the syntactic forms of the phrases are identified.

There are many different ways in which the meaning of verbs can be analysed and it is

hard to see what kind of analysis is relevant for the grammar. Consider, for example,

basic dictionary definitions of the verbs used in (2.1-2.3) given in (2.4).

(2.4)

see to notice people and things with your eyes

laugh to smile while making sounds with your voice that

show you are happy or think something is funny

tell to say something to someone, usually giving them

information

Cambridge Dictionaries Online

http://dictionary.cambridge.org/

In the definitions above, the meaning of the verbs is analysed into smaller components.

They state, for example, that seeing involves eyes, things, and people, that laughing

involves sounds, showing that you are happy, and something funny, and that telling

involves something, someone, and giving information. The units which are identified as

components of the verbs’ meaning are very different in nature: some are nouns with

specific meaning, some are pronouns with very general meaning, some are complex

phrases.

In theoretical approaches to the meaning of verbs, like in lexicography, the analysis re-

sults in identifying smaller, more primitive notions of which the meaning is composed.

Unlike lexicographic analysis, however, theoretical analysis aims at defining and organ-

ising these notions having in mind the language system as a whole, and not only the

meaning of each verb separately. This implies establishing general components which

apply across lexical items and which play a role in the rules of grammar.

35


2.1.1. The relational meaning of verbs

The most important general distinction made in the theory of lexical structure of verbs

is the one between the relational meaning and the idiosyncratic lexical content. In the

definition of the verb see given in (2.4), for example, things and people belong to the

relational structure, while eyes belong to the idiosyncratic content. In the case of the

verb laugh, all the components listed in the definition are idiosyncratic. The verb tell

has two relational components (something, someone).

The relational meaning expresses the fact that the verb relates its subject with another

entity or with a property. In this sense, verbs are analysed as logical predicates which can

take one, two, three, or more arguments. This part of their lexical structure is usually

called the predicate-argument structure. It is seen as an abstract component of meaning

present in all verbs. There are only a few possible predicate-argument structures, so that

they are typically shared by many verbs, while the idiosyncratic content characterises

each individual verb.

The predicate-argument structure is the part of the lexical representation of verbs which

determines the basic shape of clauses. In a simplified scenario, a verbal predicate which

takes two arguments forms a clause with two principal constituents, as in (2.1) and in

(2.5a). One argument in the lexical structure of a verb results in intransitive clauses

(2.2 and 2.5b) and so on. Formally, the transfer of the information from the lexicon

to syntax is handled by more general mechanisms, by projection in earlier accounts

(Chomsky 1970; Jackendoff 1977; Chomsky 1986) and feature checking in newer

proposals (Chomsky 1995; Radford 2004).

In the accounts that are based on the notion of projection, lexical items project their

relational properties into syntax by forming a specific formal structure which can then be

combined only with the structures with which it is compatible. So, for instance, a two-

argument verb will form a structure with empty positions intended for its subject and

object. In principle, these positions can only be filled by nominal structures, while other

verbal, adjectival, or adverbial structures will not be compatible with these positions.

In the feature checking account, lexical items do not form any specific structures, but

they carry their properties as features, which, by a general rule, need to match between

36


the items which are to be combined in a phrase structure. For instance, the list of

features of a two-argument verb will contain one feature requiring a subject and one

requiring an object. A verb with these features can be combined only with items which

have the matching features, that is with nominal items which bear the same features.

Characterisation of possible semantic arguments of verbs depends on the theoretical

framework adopted for an analysis, but all approaches make distinctions between at

least several kinds of arguments. The kind of meaning expressed by a verb’s argument

is usually called a semantic role. Two traditional semantic roles, agent and theme are

illustrated in (2.5).

(2.5) a. [Mary]subject/agent

stopped [the car].object/theme

b. [The car]subject/theme

stopped.

There is a certain alignment between semantic roles and syntactic functions. Agents,

for instance, tend to be realised as subjects across languages, while themes are usually

objects as in (2.5a). However, the same semantic role can be realised with different

syntactic functions, as it is the case with the theme role assigned to the car in (2.5a-

b). The phenomenon of multiple syntactic realisations of the same predicate-argument

structure is known as argument alternation. The alternation illustrated in (2.5) is called

the causative alternation, because the argument which causes the car to stop (Mary) is

present in one expression (2.5a), but not in the other (2.5b). Other well-known examples

of argument alternations include the dative alternation (2.6) and the locative alternation

(2.7).

(2.6) a. [Mary]subject/agent

told [her friend]indirect-object/recipient

[a story].object/theme

b. [Mary]subject/agent

told [a story]object/theme

[to her friend].prep-complement/recipient

37


(2.7) a. [People]subject/agent

were swarming [in the exhibition hall].prep-complement/location

b. [The exhibition hall]subject/location

was swarming [with people].prep-complement/agent

In the dative alternation, the recipient role (her friend in (2.6)) can be expressed as the

indirect object which usually takes dative case (2.6a),1 or as a prepositional complement

(2.6b). In the locative alternation, the arguments which express the location and the

agent of the situation described by the verb swap syntactic functions: the location

(exhibition hall) is the prepositional complement in (2.7a) and the subject in (2.7b). The

agent (people) is in the subject position in (2.7a) and it is the prepositional complement

in (2.7b).

The view of the predicate argument structure has evolved with the developments in

linguistic theory, from the quite intuitive notions illustrated in the examples so far to

more formal and general analyses. The main changes in the theory are reviewed in the

following sections.

2.1.2. Atomic approach to the predicate-argument structure

In the earliest approaches, the roles of the semantic arguments of verbs are regarded

as simple, atomic labels. Apart from the roles illustrated in (2.5-2.7), the set of labels

commonly includes: experiencer, instrument, source, and goal, illustrated in

(2.8-2.11).2 The atomic semantic labels of the constituents originate in the notions of

“deep cases” in Case grammar (Fillmore 1968).

These labels capture common intuitions about the relational meaning of verbs which

cannot be addressed using only the notions of syntactic functions. For example, the

meanings of the subjects in (2.5a-b), as well as the role that they play in the event

1Although the dative case is not visible in most of English phrases, including (2.6a), it can be shownthat it exists in the syntactic representation of the phrases.

2The labels patient and theme are often used as synonyms (as, for example, in (Levin and RappaportHovav 2005)). If a difference is made, patient is the participant undergoing a change of state, andtheme is the one that undergoes a change of location.

38


described by the verb stop, are rather different. Mary refers to a human being who is

actively (and possibly intentionally) taking part in the event, while the car refers to

an object which cannot have any control of what is happening. This difference cannot

be formulated without referring to the semantic argument label of the constituents. A

similar distinction is made between people and the exhibition hall in (2.7a-b).

Another important intuition which is made evident by the predicate-argument repre-

sentation is that the sentences such as (2.5a) and (2.5b) are related in the sense that

they are paraphrases of each other. The same applies for (2.6a) and (2.6b) and (2.7a)

and (2.7b). The fact that the predicate-argument structure is shared by the two para-

phrases, while their syntactic structure is different, represents the intuition that the two

sentences have approximately the same meaning, despite the different arrangements of

the constituents.

(2.8) [Mary]experiencer

enjoyed the film.

(2.9) Mary opened the door [with a card].instrument

(2.10) Mary borrowed a DVD [from the library].source

(2.11) Mary arrived [at the party].goal

Finally, the predicate-argument representation is useful in establishing the relationship

between the sentences which express the same content across languages. As the exam-

ples in (2.12) show, the relational structure of the verbs like in English and plaire in

French is the same, despite the fact that their semantic arguments have inverse syntactic

functions.

(2.12) a. [Mary]subject/experiencer

liked [the idea].object/theme

(English)

39


b. [L’idee]subject/theme

a plu [a Marie].prep-complement/experiencer

(French)

Although the predicat-argument structure proves to be a theoretically necessary level of

representation of the phrase structure, it was soon shown that the concept of semantic

roles as atomic labels for the verbs’ arguments is too naıve with respect to the reality of

the observations that it is intended to capture.

First of all, the set of roles is not definitive. There are no common criteria which

define all possible members of the set. New roles often need to be added to account for

different language facts. For example, the sentence in (2.9) can be transformed so that

instrument is the subject as in (2.13), but if we replace the card with the wind as

in (2.14), the meaning of this subject cannot be described with any of the labels listed

so far. It calls for a new role — cause or immediate cause (Levin and Rappaport

Hovav 2005). Similarly, many other sentences cannot be described with the given set

of roles. This is why different analyses keep adding new roles (such as beneficiary,

destination, path, time, measure, extent etc.) to the set.

(2.13) [The card]instrument

opened the door.

(2.14) [The wind]cause

opened the door.

Another problem posed by the atomic view of semantic roles is that there are no trans-

parent criteria or tests for identifying a particular role. Definitions of semantic roles do

not provide sets of necessary and sufficient conditions that can be used in identifying

the semantic role of a particular argument of a verb. For example, agent is usually de-

fined as the participant in an activity that deliberately performs the action, goal is the

participant toward which an action is directed,3 and source is the participant denoting

the origin of an action. These definitions, however, do not apply in many cases, as noted

by Dowty (1991). For example, both Mary and John in (2.15) seem to act voluntarily

3Dowty analyses to Mary in (2.15a) as goal, while the role of this constituent would be analysed asrecipient by other authors, which further illustrates the problem.

40


in both sentences, which means that they both bear the role of agent. Furthermore,

John is not just agent, but also source, while Mary is both agent and goal.

(2.15) (a) [John]?

sold the piano [to Mary]?

for $1000.

(b) [Mary]?

bought the piano [from John]?

for $1000.

(Dowty 1991: 556)

The example in (2.15) shows that the relational structure of such sentences cannot be

described by assigning a single and distinct semantic label to each principal constituent

of the clause. The meaning of the verbs’ arguments seems to express multiple relations

with the verbal predicate.

There is one more observation which cannot be addressed with the simple view of se-

mantic labels. This is the fact that the meaning of the roles is not equally distinct in all

the cases. Some roles obviously express similar meanings, while others are very differ-

ent. Furthermore, semantic clustering of the roles seems to be related with the kinds of

syntactic functions that the arguments have in a phrase. For example, the arguments

which are realised as subjects in (2.9), (2.13), and (2.14), agent, instrument, cause

respectively, constitute a paradigm — they can be replaced by each other in the same

context. It has been noticed that two of these roles, agent and cause, can never occur

together in the same phrase. On the other hand, the roles such as source and goal

are in a syntagmatic relation: they tend to occur together in the same phrase. The

traditional view of semantic roles as a set of atomic notions does not provide a means

to account for these facts.

Different theoretical frameworks have been developed in the linguistic literature to

deal with these problems and to provide a more adequate definitions of the predicate-

argument relations. Studying in more detail how semantic arguments of verbs are re-

alised in the phrase structure, some authors (Larson 1988; Grimshaw 1990) propose a

universal hierarchy of the arguments. The order in the hierarchy is imposed by the syn-

tactic prominence of the arguments. For example, agents are at the top of the hierarchy,

41


which means that they take the most prominent position in the sentence, the subject

position. Next in the hierarchy are themes. They are typically realised as direct objects,

but they can also be realised as subjects if agents are not present in the representation.

Lower arguments are realised as indirect objects and prepositional complements. We do

not discuss these proposals further as the view of the arguments does not significantly

depart from the atomic notions.

In the following sections, we take a closer look into the analyses which propose decom-

posing the predicate-argument structure into a set of more primitive notions. We start

with the approaches based on a decomposition of semantic roles into features or proper-

ties. Then we move to the approaches based on a decomposition of verbal meaning into

multiple predicates.

2.1.3. Decomposing semantic roles into clusters of features

An obvious direction for overcoming the problems posed by the atomic view of the

predicate-argument relationship is to decompose the notions of individual roles into

features or properties. Using a limited set of features for defining all the roles should

provide more systematic and more precise definitions of roles. It should also enable

defining a role hierarchy that can group the roles according to properties that they

share. Two approaches to the feature-based approaches to semantic roles are described

in this section.

Proto-roles

Dowty (1991) concentrates on argument selection — the principles that languages use

to determine which argument of a predicate can be expressed with which grammatical

function. Dowty (1991) argues that discrete semantic types of arguments do not exist

at all, but that the arguments are rather divided into only two conceptual clusters —

proto-agent and proto-patient. These clusters are understood as categories in the

sense of the theory of prototypes (Rosch 1973), which means that they have no clear

boundaries, and that they are not defined with sets of necessary and sufficient conditions.

42


These categories are represented with their prototypical members, with other members

belonging to the categories to a different degree. The more the members are similar to

the prototypes the more they belong to the category.

Looking into different realisations of subjects and objects and the semantic distinctions

that they express in different languages, Dowty proposes lists of features that define the

agent and the patient prototype. Each feature is illustrated by the sentence whose

number is indicated.

agent:

a. volitional involvement in the event or state (2.16)

b. sentence (and/or perception) (2.17)

c. causing an event or change of state in another participant (2.18)

d. movement (relative to the position of another participant) (2.19)

(e. exists independently of the event named by the verb) (2.20)4

patient

a. undergoes change of state (2.21)

b. incremental theme (2.22)

c. causally affected by another participant (2.23)

d. stationary relevant to movement of another participant (2.24)

(e. does not exist independently of the event, or not at all) (2.25)

(2.16) [Bill] is ignoring Mary.

(2.17) [John] sees Mary.

4Dowty uses the parentheses to express his own doubts about the relevance of the last feature in bothgroups.

43


(2.18) [Teenage unemployment] causes delinquency.

(2.19) [Water] filled the boat.

(2.20) [John] needs a new car.

(2.21) John made [a mistake].

(2.22) John filled [the glass] with water.

(2.23) Smoking causes [cancer].

(2.24) The bullet overtook [the arrow].

(2.25) John built [the house].

These examples illustrate the properties in isolation, the phrases used in contexts where

syntactic constituents are characterised with only one of the properties. Prototypical

realisations would include all agent properties for subjects and all patient properties

for objects.

These properties are conceived as entailments that are contained in verbs meaning spec-

ifying the value for the cognitive categories that people are actually concerned with:

whether an act was volitional, whether it was caused by something, whether there were

emotional reactions to it, and so on. (Dowty 1991: 575)

The relation between a verb’s meaning and its syntactic form can be formulated in the

following way: If a verb has two arguments, the one that is closer to the agent prototype

is realised as the subject, and the one that is closer to the patient prototype is realised

as the object. If there are three arguments of a verb, the one that is in between these two

ends is realised as a prepositional object. This theory can be applied to explain certain

phenomena concerning the interface between semantics and syntax. For example, the

existence of “double lexicalizations” such as those in (2.15) that are attested in many

different languages with the same types of verbs can be explained by the properties of

their arguments. Both arguments that are realised in (2.15) are agent-like arguments

(none of them being a prototypical agent), so the languages tend to provide lexical

elements (verbs) for both of them to be realised as subjects.

44


Dowty’s theory provides an elaborate framework for distinguishing between verbs’ argu-

ments, accounting for numerous different instances of arguments. These characteristics

make it a suitable conceptual framework for large-scale data analysis. Recently, Dowty’s

notions have been used as argument descriptors in a large-scale empirical study of mor-

phosyntactic marking of argument relations across a wide range of languages (Bickel

et al. To appear), as well as in a large-scale annotation project (Palmer et al. 2005a).

Dowty’s approach, however, does not address issues related to syntax, such as different

syntactic realisations of the same arguments. The approach reviewed in the following

subsection is more concentrated on these issues.

The Theta System

Unlike Dowty, who assumes monostratal syntactic structure of phrases (Levin and Rap-

paport Hovav 2005), Reinhart (2002) sets the discussion on semantic roles in the context

of derivational syntax. The account proposed by Reinhart (2002) offers an elaborate

view of the interface between lexical representation of verbs and syntactic derivations.

It assumes three independent cognitive modules — the systems of concepts, the compu-

tational system (syntax), and the semantic inference systems. Linguistic information is

first processed in the systems of concepts, then passed on to the computational system,

and then to the semantic inference systems. The theta system5 belongs to the systems

of concepts. It enables the interface between all the three modules. It consists of three

parts:

a. Lexical entries, where theta-relations of verbs are defined.

b. A set of arity operations on lexical entries, where argument alternations are pro-

duced.

c. Marking procedures, which finally shape a verb entry for syntactic derivations.

5In the generative grammar theoretical framework, semantic roles are often referred to as thematicroles or, sometimes, as Θ-roles, indicating that the meaning expressed by the semantic argumentsof verbs is not as specific as the traditional labels would suggest.

45


There are eight possible theta relations that can be defined for a verb and that can

be encoded in its lexical entry. They represent different combinations of values for two

binary features: cause change (feature [c]) and mental state (feature [m]). They

can be related to the traditional semantic role labels in the following way:

a) [+c+m] — agent

b) [+c−m] — instrument (. . .)

c) [−c+m] — experiencer

d) [−c−m] — theme / patient

e) [+c] — cause (unspecified for / m); consistent with either (a) or (b)

f) [+m] — ? (Candidates for this feature-cluster are the subjects of verbs like love,

know, believe)

g) [−m] — (unspecified for / c): subject matter / locative source

h) [−c] — (unspecified for / m): roles like goal, benefactor; typically dative (or

PP).

The verb entries in the lexicon can be basic or derived. There are three operations that

can be applied to the basic entries resulting in derived entries: saturation, reduction,

and expansion.

Saturation is applied to the entries that are intended for deriving passive constructions.

It defines that one of the arguments is just an existential quantification and that it is

not realised in syntax. It is formalised as described in the example of the verb wash:

a) Basic entry: wash(θ1, θ2)

b) Saturation: ∃x(wash(x, θ2))

c) Max was washed : ∃x(x washed Max)

46


Reduction can apply in two ways. If it applies to the argument that is realised within

the verb phrase in syntax (typically, the direct object) it reduces the verb’s argument

array to only one argument, so that the meaning of the verb is still interpreted as a

two-place relation, but as its reflexive instance ((2.26b) vs. (2.26a)). If it applies to the

argument that is realised outside of the verb phrase, which means as the subject in a

sentence, it eliminates this argument from the array of verb’s arguments completely, so

that the verb is interpreted as a one-place relation ((2.26c) vs. (2.26a)).

(2.26) (a) Mary stopped the car.

(b) Mary stopped.

(c) The car stopped.

Expansion is an operation usually known as causativization. It adds one argument —

agent — to the array of the verb (2.27b vs. 2.27a).

(2.27) (a) The dog walked slowly.

(b) Mary walked the dog slowly.

All these operations take place in the lexicon, producing different outputs. While the op-

erations of saturation and reduction produce new variations of the same lexical concept,

expansion creates a whole new concept.

Before entering syntactic derivations, the concepts undergo one more procedure, the

marking procedure, which assigns indices to the arguments of verbs. These indices serve

as a message to the computational system as to where to insert each argument in the

phrase structure. Taking into consideration the number and the type of the feature

clusters that are found in a verb entry, they are assigned according to the following

rules:

Given an n-place verb entry, n > 1,6

6Insertion of a single argument as subject follows from a more general syntactic rule, namely theExtended Projection Principle, which states that each clause must have a subject.

47


a) Mark a [−] cluster with index 2.

b) Mark a [+] cluster with index 1.

c) If the entry includes both a [+] cluster and a fully specified cluster [/α, /−c],7

mark the verb with the ACC feature.8

This marking is associated with the following instructions for the computational sys-

tem:

a) When nothing rules this out, merge externally.

b) An argument realising a cluster marked 2 merges internally.

c) An argument with a cluster marked 1 merges externally.

The operation of internal merging joins a new constituent to an existing structure within

a verb phrase, while the operation of external merging inserts a new constituent in the

existing structure of a sentence outside of the phrase headed by the verb in question.

The result of an internal merge is usually a syntactic relation between a verb and its

object, while external merge forms a relation between a verb and its subject.

With this system, some generalizations concerning the relation between theta roles and

syntactic functions can be stated. Arguments that realise [+] clusters ([+m−c] agent,

[+c] cause, [+m] ?) are subjects. Since there can be only one subject in a sentence,

they exclude each other. Arguments that realise [−] clusters ([−m−c] patient, [−m]

subject matter, [−c] goal) are objects. Only the fully specified one can be the direct

object (introducing the ACC feature to the verb). The others (underspecified ones) have

to be marked with a preposition or an inherent case (e.g. dative), thus realised as indirect

objects.

7A cluster that is specified for both features, where one of them has to be [−c] and the other can beany of the following: [+m], [−m], [+c], [−c].

8ACC stands for accusative. This feature determines whether a verb assign accusative case to itscomplement.

48


Arguments that are specified for both features, but with opposite values ([+m−c] expe-

riencer and [−m+c] instrument) are neutral. They have no indices, so they can be

inserted into any position in the phrase structure that is available at the moment of their

insertion. The same applies to the arguments that are encoded as the only argument of

a verb.

Summary

The approaches of Dowty (1991) and Reinhart (2002) reviewed in this section deal with

issues in the traditional predicate-argument analysis by proposing small sets of primitive

notions which can be seen as semantic components of the traditional argument labels.

With such decomposition, the similarities between some arguments (e.g. agent, cause,

and instrument), as well as the constraints in their realisations in phrases (e.g. if a

clause expresses an agent of an event, it cannot express another distinct cause of

the same event), follow from the features that characterise them. The generalisations

proposed as part of these approaches contribute to a better understanding of how the

interface between lexicon and syntax operates, capturing a wide range of observations.

However, the sets of features used in these accounts are not motivated by some other

more general principles. In other words, these accounts do not address the issue of why

we have exactly the sets of features proposed in the two theories and not some others.

The approaches to the predicate-argument structure reviewed in the following section

propose deeper semantic analysis exploring the origins of the argument types.

2.1.4. Decomposing the meaning of verbs into multiple predicates

In theories of predicate decomposition, it is assumed that verbs do not describe one

single relation, but more than one. These relations are regarded as different components

of an event described by the meaning of a verb, often referred to as subevents. Some

of these components can be rather general and shared by many verbs, while others

are idiosyncratic and characteristic of a particular verb entry. In this framework, each

predicate included in the lexical representation of a verb assigns semantic roles to its

49


arguments. The syntactic layout of a clause depends, thus, on the number and the

nature of the predicates of which the meaning of its heading verb is composed.

Many approaches to predicate decomposition are influenced by the work of Hale and

Keyser (1993) who are the first to prpose a formal syntactic account the relational

meaning of verbs. Hale and Keyser (1993) propose a separate level of lexical repre-

sentation of verbs — lexical relational structure (LRS). The components of conceptual,

idiosyncratic meaning of a verb are the arguments of the relations grouped in the LRS.

The relational part of lexical representation, for example, is the same for the verbs get in

(2.28) and bottle in (2.29) indicating that machine did something to the wine. The dif-

ference in meaning between these two verbs is explained by two different incorporations

of idiosyncratic components in the relational structure. In the first case, the relational

structure incorporates the verb get with its own complex structure, while in the second

case, it incorporates the noun bottle.

(2.28) A machine got the wine into bottles.

(2.29) A machine bottled the wine.

Different approaches which follow this kind of analysis offer different representations of

the relational structure, depending on what organizational principle is taken as a basis

for event decomposition.

Aspectual event analysis

Aspectual analysis of events takes into consideration temporal properties of verbs’ mean-

ing. More precisely, it decomposes the relational part of lexical representation of verbs

into a number of predicates which correspond to the stages in the temporal development

of the event. These predicates take arguments which are then realised as the principal

constituents of clauses. As an illustration of the phenomena that are of interest for the

aspectual decomposition of verbs’ meaning, consider the sentences in (2.30-2.31).

(2.30) a. Mary drank [a bottle of wine] in 2 hours / ? for two hours.

b. Mary drank [wine] for 2 hours / * in two hours.

50


(2.31) a. Mary crammed [the pencils] into the jar.

b. Mary crammed [the jar] with the pencils.

The examples in (2.30) show that the choice of the adverbial with which the verb drink

can be combined (in two hours vs. for two hours) depends on the presence or absence

of the noun bottle in the object of the verb. Intuitively, we know that the adverbials

such as in two hours are compatible only with the events which are understood as

completed. The event expressed in the sentence in (2.30a), for example, is completed

because the sentence implies that there is no more wine in the bottle. What makes this

event completed and, thus, compatible with the adverbial in two hours is precisely the

presence of the noun bottle in the object. This noun quantifies the substance which is

the object (wine) and, at the same time, it quantifies the whole event which includes

it. This fact points to the presence of a predicate in the relational structure of the verb

which takes the noun bottle as its argument. This predicate relates the other parts of the

lexical structure of the verb (or of the event described by the verb) with an end point.

The nature of the end point is specified by the argument of this predicate, that is by the

argument which is realised as the direct object in the phrase. If the direct object does

not provide the quantification, the whole event is interpreted as not quantified, as in

(2.30b). The verb in (2.30b) is not compatible with the adverbial in two hours because

its object is not quantified.

The examples in (2.31) show how an alternation of the arguments of a verb can change

the temporal interpretation of the event described by the verb. The event in (2.31a)

lasts until all the pencils are in the jar, while the event in (2.31b) lasts until the jar is

full. The argument of the temporal delimitation predicate, which is usually realised as

the direct object in a phrase, is known as the incremental theme (Krifka 1998). The

adjective incremental in the term refers to the fact that the theme argument of the verb

changes incrementally in the course of the event described by the verb. The degree of

change in the theme “measures out” the development of the event.

An influential general approach to aspectual decomposition of verbal predicates is pro-

posed by Ramchand (2008). Looking at the event as a whole, Ramchand (2008) proposes

several predicates which take arguments such as initiator, undergoer, path, and

resultee. The predicates represent the subevents of the event described by a verb:

51


the predicate whose argument is initiator represents the beginning of the event, those

that take the arguments undegoer and path are in the middle, and the one whose

argument is resultee is in the end. These predicates are added to the representation in

the course of syntactic derivations, but only if this is allowed by the lexical specifications

of verbs.

In an analysis of the example (2.31a) in this framework, Mary is initiator, the pencils

are both undergoer and resultee, and the jar is another argument of the last (re-

sulting) predicate. In (2.31b) on the other hand, the argument of both the middle and

the end predicate is the jar.

This view of the semantic structure relates the complexity of the event described by the

verb with the complexity of its argument structure, and by this, with the complexity of

the structure of the clause that is formed with the verb. It should be noted, however,

that this decomposition does not address all the temporal properties of phrases, but

only those which are implicit to the meaning of the verbs.

Causal event analysis

In a causal analysis, events are analysed into entities (participants) that interact accord-

ing to a particular energy flow or force-dynamic schemata (Talmy 2000). The main

concept in this framework is the direction that the force can take. Semantic properties

of verbs such as aspect and argument structure are seen as a consequence of a particular

energy flow involved in the event described by the verb. If a verb describes an event

were some energy is applied, it will be an action verb, otherwise it will describe a state.

In the domain of argument structure, the participant that is the source of energy will

have the role of agent and the one that is the sink of the energy will have the role of

patient.

This approach has been applied to account for different senses of the same verb, as well

as for its different syntactic realisations. Both different verb senses and their argument

alternations are explained by the shifts in the energy flow with the idiosyncratic meaning

that stays unchanged. The following examples illustrate different force-dynamic patterns

for the verb take.

52


(2.32) Sandy took the book { from Ashley / off the table }.

(2.33) (a) Sandy took the book to Ashley.

(b) Sandy took Ashley (to the movies).

The action in (2.32) is self-oriented, with Sandy being the energy source and sink at the

same time. In (2.33), another participant (Ashley / to the movies) is more important,

since it indicates the direction of energy. These are two different senses of the verb take.

The difference is reflected in the fact that this argument can be omitted in sentences

like (2.32) whenever it is indefinite or indeterminate, while in sentences like (2.33), it

can only be omitted if it can be interpreted from the context.

2.1.5. Summary

In the approaches to the analysis of the predicate-argument structure outlined so far, the

meaning of verbs and their arguments is described with relatively small inventories of

descriptive notions. The proposed accounts decompose the predicate-argument structure

into more primitive notions in an attempt to reduce the number of theoretical notions

to a minimum. The Theta system by Reinhart (2002), for example, accounts for a

wide range of linguistic facts using only three notions: mental state, cause change,

and presence/absence. Similarly, the aspectual decomposition proposed by Ramchand

(2008) results in only four principal components which play a role in accounting for

many different argument realisations and interpretations.

Theoretical accounts reviewed in this section identify important generalisations about

the lexical representation of verbs. The generalisations are, however, not tested on a

wide range of verbs, but mostly on a small set of examples either provided by the authors

of the proposals themselves or taken from a common set of examples frequently cited in

the literature. Applying theoretical generalisations to a larger set of verb instances in a

more practical analysis is not straightforward. Some approaches to large-scale analysis

of the predicate-argument structure are discussed in the following section.

53


2.2. Verb classes and specialised lexicons

We have assumed so far that the predicate-argument relational structure is the part of

lexical representation which is shared by different verbs, while the idiosyncratic lexical

content is specific to each individual verb. However, if we take a closer look at the

inventory of verbs in a language, this distinction turns out to be a simplified view of the

organisation the inventory. We intuitively group together not only the verbs with the

same predicate-argument structure, but also the verbs with similar lexical content. Such

groups are, for example, verbs of motion (e.g. come, go, fall, rise, enter, exit, pass),

state verbs (e.g. want, need, belong, know), verbs of perception (e.g. see, watch, notice,

hear, listen, feel) etc.

2.2.1. Syntactic approach to verb classification

The members of semantic classes tend to be associated with the same types of syntactic

structures. For example, verbs of motion are usually intransitive, verbs of perception are

usually transitive, while the verbs that describe states can be associated with a variety

of different structures. However, it has been noticed that verbs which belong to the

same semantic class do not always participate in the same argument alternations. For

example, the verbs bake in (2.34) and make in (2.35) have similar meanings in that they

are both verbs of creation and that they can take the same kind of objects.

(2.34) a. Mary baked a cake.

b. The cake baked for 40 minutes.

(2.35) a. Mary made a cake.

b. *The cake made for 40 minutes.

Despite the obvious parallelism, the verb bake participates in the causative alternation

(2.34b) is grammatical), while the verb make does not ((2.35b) is not grammatical).

This contrast suggests that the two verbs have different lexical representations. There

54


should be an element which is present in the structure of one verb and missing in the

structure of the other, causing the difference in the syntactic patterns.

On the basis of this assumption Levin (1993) studies possible argument realisations of

a large number of verbs and proposes a comprehensive classification which combines

semantic and syntactic groupings. The aim of Levin’s analysis is to use the information

about argument alternations as behavioural, observable indicators of the components of

verbs’ meaning which are grammatically relevant.

Having pointed out many constraints and distinctions which call for theoretical accounts,

Levin’s classification has been often referred to in the subsequent work on the predicate-

argument structure, including Levin’s own work (Levin and Rappaport Hovav 1994;

1995). An example of a phenomenon identified in Levin’s classification which has re-

ceived a proper theoretical account is the distinction between verbs such as freeze, melt,

grow, which are known as unaccusatives, and emission verbs such as glow, shine, beam,

sparkle. The two groups are similar in that they consist of intransitive verbs which

take the same kind of arguments — non-agentive, non-volitional — as subjects. The

syntactic properties of the two groups, however, are different. While unaccusatives par-

ticipate in the causative alternation, the emission verbs do not, which groups them with

semantically very different agentive intransitive verbs such as walk, run, march, gallop,

hurry. Reinhart (2002) employs the notions developed in the framework of the Theta

System (see Section 2.1.3) to explain this fact by different derivations of unaccusatives

and emission verbs. Unaccusatives are derived lexical entries (derived from transitive

verbs). Their argument is marked with the index 2, as the internal argument of the

transitive verb. By the operation of reduction, the other argument is removed. The

remaining argument is merged internally, even if it stays the only argument of the verb

due to the fact that is marked with the index 2. It then moves to the position of the

subject to satisfy general syntactic conditions. As for emission verbs, their subject is

originally the only argument. This is why it cannot be marked. And since it is not

marked, it is merged to the first position available — and this is the external position

of the subject.

The systematic analysis of a large set of verbs proposed by Levin (1993) proved to be es-

pecially important for the subsequent empirical approaches to the meaning of verbs. The

55


classification has often been cited as the reference resource for selecting specific groups

of verbs for various purposes, including the experiments presented in this dissertation.

The more recent work on the argument alternation is concentrated on the conditions

which determine different syntactic realisations of verbs’ arguments in alternations.

Beavers (2006) revisits a range of alternations, especially those which involve arguments

switching between the direct object and a prepositional complement, arguing that gen-

eral semantic relationships between syntactic constituents directly influence their posi-

tion in a clause, and not only the relationship of the arguments with verbal predicates.

Beavers (2006) proposes a set of semantic hierarchies along different dimensions, such

as the one illustrated in (2.36): the higher the interpretation of an argument in the

hierarchy, the more distant its syntactic realisation from the direct object.

(2.36) Affectedness scale:

PARTICIPANT ⊂ IMPINGED ⊂ AFFECTED

⊂ TOTALLY AFFECTED

Bresnan (2007) takes an empirical approach proposing a statistical model of speakers’

choice between two options provided by the dative alternation. The study first shows

that human judgement of acceptability of syntactic constructions are influenced by the

frequency of the constructions. It then shows that several factors are good predictors

of human judgements. If the recipient role is characterised as nominal, non-given, in-

definite, inanimate, and not local in the given spacial context, it is likely to be realised

as a prepositional complement, while it is characterised with the opposite features (as

pronominal, given, definite, animate, and local) it is likely to be realised as the indirect

object. Bresnan and Nikitina (2009) offer an explanation of the speaker’s choice based

on the interaction of two opposed tendencies. On the one hand, there is the tendency

of semantic arguments to be aligned with the syntactic functions, more prominent argu-

ments are realised as more prominent syntactic functions (like the direct and the indirect

object), while less prominent arguments are realised as prepositional complements. On

the other hand, the form of the prepositional phrase expresses the relationship between

the verb and its complement in a more transparent way. Hence, if the argument is seman-

tically prominent enough, it will be assigned a less transparent, but more syntactically

prominent functions. Otherwise, it will be realised as a prepositional complement.

56


The more recent developments in the approach to argument alternations, however, have

not been followed by a large-scale implementation. The comprehensive resources which

have been developed up to the present day do not make reference to these generalisa-

tions.

2.2.2. Manually annotated lexical resources

Three big projects are concerned with providing extensive descriptions of the predicate-

argument relations for English words. They are described in the following subsections.

We start by describing FrameNet. As our detailed review shows, this resource imple-

ments the least theoretical view of the predicate argument structure based on the atomic

semantic role analysis. Nevertheless, this resource is the one that is most frequently used

as a reference for developing similar resources for other languages. The second resource,

PropBank, implements Dowty (1991)’s view of the predicate-argument structure, but

with significant simplifications which brings the implementation closer to the atomic

view. This resource has been frequently used for machine learning experiments due to

the fact that, in addition to the lexicon of verbs, it provides a large corpus of texts

manually annotated with the proposed predicate-argument analysis. The last resource

that we discuss is VerbNet. Although it implements Levin (1993)’s classification, this

resource too relies on the atomic view of semantic roles, assigning traditional semantic

role labels to the arguments of verbs.

FrameNet

FrameNet is an electronic resource and a framework for explicit description of the lexical

semantics of words. It is intended to be used by lexicographers, but also by systems for

natural languages processing (Baker et al. 1998). It consists of three interrelated

databases:

a. Frame database, the core component describing semantic frames that can be expressed

by lexical units.

57


b. Annotated example sentences extracted from the British National Corpus (Burnard

2007) with manually annotated frames and frame elements that are described in the

Frame database.

c. Lexicon, a list of lexical units described in terms of short dictionary definitions and

detailed morpho-syntactic specifications of the units that can realise their arguments in

a sentence.

The frame database (1 179 frames)9 contains descriptions of frames or scenes that can

be described by predicating lexical units (Fillmore 1982), such as verbs, adjectives,

prepositions, nouns. Each scene involves one or more participants. The predicating units

are referred to as “targets”, and the participant which are combined with predicating

units as “frame elements”. One frame can be realised in its “core” version including

“core” frame elements, or it can be realised as a particular variation, including additional

frame elements that are specific for the variation. For example, the target unit for the

frame Accomplishment can be one of the verbs accomplish, achieve, bring about, or one of

the nouns accomplishment, achievement. The core frame elements for this frame are:

a. agent: The conscious entity, generally a person, that performs the intentional act

that fulfills the goal.

b. goal: The state or action that the agent has wished to participate in.

The definition of the frame itself specifies the interaction between the core frame ele-

ments:

After a period in which the agent has been working on a goal, the agent

manages to attain it. The goal may be a desired state, or be conceptualised

as an event.

For the non-core realisations, only the additional frame elements are defined.

The frames are organised into a network by means of one or more frame-to-frame rela-

tions that are defined as attributes of frames. (Note thought that not all frames could

9Online documentation at https://framenet.icsi.berkeley.edu/fndrupal/current_status, ac-cessed on 5 June 2014.

58


be related to other frames.) Defining the relations enables grouping related frames ac-

cording to different criteria, so that the annotation can be used with different levels of

granularity. There are six types of relations that can be defined:

• Inherits From / Is Inherited By: relates an abstract to a more specified frame with

the same meaning, e.g. Activity (no lexical units) inherits from Process (lexical

unit process.n) and it is inherited by Apply heat (lexical units bake.v, barbecue.v,

boil.v, cook.v, fry.v, grill.v, roast.v, toast.v, ...).

• Subframe of / Has Subframes: if an event described by a frame can be divided

into smaller parts, this relation holds between the frames that describe the parts

and the one that describes the whole event, e.g. Activity has subframes: Activity

abandoned state, Activity done state (lexical units done.a, finished.a, through.a),

Activity finish (lexical units complete.v, completion.n, ...), Activity ongoing (lexical

units carry on.v, continue.v, keep on.v, ...), Activity pause (lexical units freeze.n,

freeze.v, pause.n, take break.v, ...), Activity paused state, Activity prepare, Activ-

ity ready state (lexical units prepared.a, ready.a, set.a), Activity resume (lexical

units renew.v, restart.v, resume.v), Activity start (lexical units begin.v, beginner.n,

commence.v, enter.v, initiate.v, ...), Activity stop (lexical units quit.v, stop.v, ter-

minate.v, ...).

• Precedes / Is Preceded by: holds between the frames that describe different parts of

the same event, e.g. Activity pause precedes Activity paused state and is preceded

by Activity ongoing.

• Uses / Is Used By: connects the frames that share some elements, e.g. Accom-

plishment (lexical units accomplish.v, accomplishment.n, achieve.v, achievement.n,

bring about.v) uses Intentionally act (lexical units act.n, act.v, action.n, activity.n,

carry out.v,, ...).

• Perspective on / Is perspectivised in: holds between the frames that express differ-

ent perspectives on the same event, e.g. Giving (lexical units gift.n, gift.v, give.v,

give out.v, hand in.v, hand.v, ...) is a perspective on Transfer (lexical units trans-

fer.n, transfer.v), Transfer can also be perspectivised in Receiving (lexical units

accept.v, receipt.n, receive.v).

59


• Is Causative of: e.g. Apply heat is causative of Absorb heat (lexical units bake.v,

boil.v, cook.v, fry.v,...)

The relations of inheritance, using, subframe, and perspective connect specific frames

to the corresponding more general frames, but in different ways. The specific frame is

a kind of the general frame in inheritance. Only a part of the specific frame is a kind

of the general frame in using. A subframe is a part of another frame. The other two

relations do not involve abstraction, they hold between the frames of the same level of

specificity.

Frames and frame elements can also be classified into semantic types that are not based

on the hierarchies described above, but that correspond to some ontologies that are

commonly referred to (such as WordNet (Fellbaum 1998)). For example, frames are

divided into non-lexical (e.g. Activity) and lexical (e.g. Accomplishment). Similarly, the

frame element agent belongs to the type “sentient”, and theme belongs to the type

“physical object”.

Annotated examples such as (2.37-2.41) are provided for most of the frame versions.

(2.37) [Iraq]agent

had [achieved]target

[its programme objective of producing nuclear weapons].goal

(2.38) Perhaps [you]agent

[achieved]target

[perfection]goal

[too quickly].manner

(2.39) [He]agent

has [only partially]degree

[achieved]target

[his objective].goal

(2.40) [These positive aspects of the Michigan law]goal

may, however, have been

[achieved]target

at the expense of simplicity. [CNI]agent

60


Frame element Syntactic realization (phrase types and their functions)agent CNI.– NP.Ext PP[by].Dep PP[for].Depcircumstances PP[in].Dep PP[despite].Dep PP[as].Dep AJP.Dep

PP[on].Depdegree AVP.Depexplanation PP[since].Dep PP[because of].Depgoal NP.Ext NP.Obj NP.Dep PP[in].Depinstrument PP[with].Dep NP.Ext VPing.Depmanner AVP.Dep PP[in].Depmeans PP[through].Dep PP[by].Dep PP[in].Dep NP.Extoutcome PP[at].Dep NP.Extplace PP[in].Dep PP[at].Deptime Sfin.Dep AVP.Dep PP[in].Dep PP[at].Dep

PP[after].Dep

Table 2.1.: Frame elements for the verb achieve

(2.41)A programme of national assessment began in May 1978 and concerned itself with

[the standard]goal

[achieved]target

[by 11 year olds].agent

The units of the lexicon are word senses (12 754 units). The entries contain a short

lexical definition of the sense of the word, the frame that the unit realises, as well as the

list of the frame elements that can occur with it. They also contain two more pieces of

information on frame elements: the specification of the syntactic form that each frame

element can take and a list of possible combinations of the frame elements in a sentence.

For example, the verb achieve realises the frame Accomplishment. The frame elements

that can occur with it are listed in Table 2.1.

The first row in Table 2.1 states that the frame element agent occurs with the verb

achieve and that it can be realised as Constructional null instantiation (CNI), which

is most often the case in passive sentences (2.40), or as a noun phrase external to the

61


Agent GoalNP NPExt ObjAgent Goal MannerNP NP AVPExt Obj DepAgent Degree GoalNP AVP GOALExt Dep Obj

Table 2.2.: Some combinations of frame elements for the verb achieve.

verb phrase headed by the target verb, which is most often the subject of a sentence

(2.37-2.39), or as a prepositional phrase headed by the preposition by and realizing

the grammatical function of dependent,10 or as a prepositional phrase headed by the

preposition for with the same grammatical function. Possible syntactic realizations for

the other frame elements are described in the same way.

Since not all frame elements can be combined with all the others in a sentence, the

possible combinations of the frame elements are also listed. Some of the possible combi-

nations for the verb achieve, those that correspond to the examples (2.37-2.39), are listed

in Table 2.2. The original entry for this verb contains 19 combinations in total. Each

of the combinations can have several versions depending on the type and grammatical

function of the constituents that realise the frame elements.

The Proposition Bank (PropBank)

PropBank is a resource which consists of an annotated corpus of naturally occurring

sentences and a lexicon of verbal predicates with explicitly listed possible arguments. It

is intended to be used for developing systems for natural language understanding that

10In the system of grammatical functions used in FrameNet, the standard distinction between a comple-ment and a modifier is not made. They are both considered dependent constituents — dependents.(Ruppenhofer et al. 2005)

62


depend on semantic parsing, but also for quantitative analysis of syntactic alternations

and transformations.

The corpus contains 2 499 articles (1 million words) published in the Wall Street Journal

for which the syntactic structure was annotated in the The Penn Treebank Project (Mar-

cus et al. 1994). The semantic roles were annotated in the PropBank project (Palmer

et al. 2005a). The labels for the semantic roles were attached to the corresponding

nodes in the syntactic trees. A simplified example of an annotated sentence is given in

(2.42). The added semantic annotation is placed between the “@” characters.

(2.42)

S

VP

VP

PP-TMP@ARGM-TMPpay01@

NN

year

JJ

last

NP@ARG1pay01@

RB

little

RB

very

VBN@rel-pay01@

paid

VBZ

has

NP-SBJ@ARG0pay01@

NN

nation

DT

The

Only a limited set of labels was used for annotation. Verbs are marked with the label

rel for relation and the participants in the situation described by the verb are marked

with the labels arg0 to arg5 for the verb’s arguments and with arg-m for adjuncts.

The numbered labels represent semantic roles of a very general kind. The labels arg0

and arg1 have approximately the same value with all verbs. They are used to mark

instances of proto-agent (arg0) and proto-patient (arg1) roles (see 2.1.3). The

value of other indices varies across verbs. It depends on the meaning of the verb, on the

type of the constituent that they are attached to, and on the number of roles present

in a particular sentence. arg3, for example, can mark the purpose, or it can mark

a direction or some other role with other verbs. The indices are assigned according

to the roles’ prominence in the sentence. More prominent are the roles that are more

closely related to the verb.

63


The arg-m can have different versions depending on the semantic type of the con-

stituent: loc denoting location, cau for cause, ext for extent, tmp for time, dis for

discourse connectives, pnc for purpose, adv forgeneral-purpose, mnr for manner, dir

for direction, neg for negation marker, and mod for modal verb. The last three labels

do not correspond to adjuncts, but they are added to the set of labels for semantic an-

notation nevertheless, so that all the constituents that surround the verb could have a

semantic label (Palmer et al. 2005a). The labels for adjuncts are more specific than the

labels for arguments. They do not depend on the presence of other roles in the sentence.

They are mapped directly from the syntactic annotation.

For example, the verb pay in (2.42) assigns two semantic roles to its arguments and one

to an adjunct. arg0 is attached to the noun phrase that is the subject of the sentence

(NP-SUBJ: The nation) and it represents the (proto-)agent. arg1 is attached to

the direct object (NP: very little). The label for the adjunct (PP-TMP: last year),

arg-m-tmp, is mapped from the syntactic label for the corresponding phrase.

The annotated corpus is accompanied with a lexicon that specifies the interpretation of

the roles for each verb in its different senses. The unit of the lexicon is a lemma (3300

verbs) containing one or more lexemes (4 500 verb senses). The interpretations for the

numbered roles are given for each lexeme separately. Table 2.3 illustrates the lexical

entry for the verb pay.

Possible syntactic realisations of the roles are not explicitly described as in FrameNet, but

they are illustrated with a number of annotated sentences, each representing a different

syntactic realisation of the role. These sentences are mostly drawn from the corpus. For

some syntactic realizations that are not attested in the corpus, example sentences are

constructed.

VerbNet

VerbNet is a database which is primarily concerned with classification of English verbs.

The approach to classification is based on the framework proposed by Levin (1993). It

takes into account two properties: a) the lexical meaning of a verb and b) the kind of

64


pay.01 Arg0: payer or buyerArg1: money or attentionArg2: person being paid, destination of attentionArg3: commodity, paid for what

pay.02 Arg0: payer(pay off) Arg1: debt

Arg2: owed to whom, person paidpay.03 Arg0: payer or buyer(pay out) Arg1: money or attention

Arg2: person being paid, destination of attentionArg3: commodity, paid for what

pay.04 Arg1: thing succeeding or working outpay.05 Arg1: thing succeeding or working out(pay off)pay.06 Arg0: payer(pay down) Arg1: debt

Table 2.3.: The PropBank lexicon entry for the verb pay.

argument alternations that can be observed in the sentences formed with a particular

verb (see Section 2.2.1 for more details).

The unit of classification in VerbNet is a verb sense. It currently covers 6 340 verb senses.

The classification is partially hierarchical, including 237 top-level classes with only three

more levels of subdivision (Kipper Schuler 2005). Each class entry includes:

• Member verbs.

• Semantic roles — All the verbs in the class assign the same roles. These roles are

semantic roles that are more general than frame elements in FrameNet, but more

specific than the numbered roles in PropBank. The label for a role in VerbNet does

not depend on context, as in FrameNet and PropBank. There is a fixed set of roles

that have the same interpretation with all verbs. Although the set of roles is fixed

in principle, its members are revised in the course of the resource development

(Bonial et al. 2011). Initially the set included the following roles:

65


Members accept, discourage, encourage, understandRoles agent [+animate | +organization] , propositionFrames HOW-S:

Example: “I accept how you do it.”Syntax: Agent V Proposition <+how-extract>Semantics: approve(during(E), Agent, Proposition)

Table 2.4.: The VerbNet entry for the class Approve-77.

Original Role set in VerbNet (23): actor, agent, asset, attribute,

beneficiary, cause, location, destination, source, experi-

encer, extent, instrument, material, product, patient, pred-

icate, recipient, stimulus, theme, time, topic.

The set was later revised to include the following roles:

Updated Role set in VerbNet (33): actor, agent, asset, attribute,

beneficiary, cause, co-agent, co-patient, co-theme, destina-

tion, duration, experiencer, extent, final time, frequency,

goal, initial location, initial time, instrument, location, ma-

terial, participant, patient, pivot, place, product, recipi-

ent, result, source, stimulus, theme, trajectory, topic.

The revised set is accompanied by a hierarchy, where all the roles are classified

into four categories: actor, undergoer, time, place.

• Selectional restrictions — defining characteristics of possible verbs’ arguments,

such as [+animate | +organization] for the role agent in Table 2.4. They can be

compared with the semantic types in FrameNet.

• Frames — containing a description of syntactic realizations of the arguments and

some additional semantic features of verbs. (Note that the VerbNet frames are

different from the FrameNet frames.) In the example class entry given in Table

2.4, only one (HOW-S) of the 5 frames that are defined in the original entry is

66


included, since the other frames are defined in the same way. The semantics of the

verb describes the temporal analysis of the verbal predicates (see Section 2.1.4).

The VerbNet database contains also information about the correspondence between the

classes of verbs and lexical entries in other resources. 5 649 links with the PropBank

lexicon entries have been specified, as well as 4 186 with the FrameNet entries.

No annotated example sentences are provided directly by the resource. However, natu-

rally occurring sentences with annotated VerbNet semantic roles can be found in another

resource, SemLink (Loper et al. 2007), which maps the PropBank annotation to the

VerbNet descriptions. Each numbered semantic role annotated in the PropBank corpus

is also annotated with the corresponding mnemonic role from the set of roles used in

VerbNet. This resource enables comparison between the two annotations and exploration

of their usefulness for the systems for automatic semantic role labelling.

Comparing the resources

The three resources that are described in the previous subsections all provide informa-

tion on how predicating words combine with other constituents in a sentences: what

kind of constituents they combine with and what interpretation they impose on these

constituents. They are all intended to be used for training systems for automatic seman-

tic parsing. However, there are considerable differences in the data provided by these

resources.

The overlap between the sets of lexical items covered by the three resources is rather

limited. For example, 25.5% of the word instances in PropBank are not covered by

VerbNet (Loper et al. 2007)), despite the fact that VerbNet contains more entries

than PropBank (3 300 verbs in PropBank vs. 3 600 verbs in VerbNet). The coverage

issue has been pointed out in the case of FrameNet too. Burchardt et al. (2009) use

English FrameNet to annotate a corpus of German sentences manually. They find that

the vast majority of frames can be applied to German directly. However, around one

third of the verb senses identified in the German corpus were not covered by FrameNet.

Also, a number of German verbs were found to be underspecified. Contrary to this,

Monachesi et al. (2007) use PropBank labels for semi-automatic annotation of a corpus

67


of Dutch sentences. Although not all Dutch verbs could be translated to an equivalent

verb sense in English, these cases were assessed as relatively rare. Samardzic et al. (2010)

and van der Plas et al. (2010) use PropBank labels to annotate manually a corpus of

French sentences. The coverage reported in these studies is around 95%. Potential

explanation of the coverage differences lies in the fact that PropBank is the only of the

three resources which is based on text samples. While the criteria for including lexical

items in the lexicons are not clear in the other two resources, the verbs and verb senses

included in PropBank are those that are found in the corpus which is taken as a starting

point.

The criteria for distinguishing verb senses are differently defined, which means that

different senses are described even for the same words. It can be noted, for example,

in Table 2.3 that PropBank introduces a new verb sense for a phrasal verb even if its

other properties are identical to those of the corresponding simplex verb, which is not

the case in the other two databases.

The information that is included in the lexicon entries is also different. We can see, for

example, that the morpho-syntactic properties of the constituents that combine with

the predicating words are described in different way. While FrameNet provides detailed

specifications (Table 2.1), VerbNet defines these properties only for some argument re-

alizations (Table 2.4). PropBank does not contain this information in the lexicon at all,

but all the instances of the roles are attached to nodes in syntactic trees in the annotated

corpus.

Finally, different sets of roles are used in the descriptions. FrameNet uses many different

role labels that depend on which frame they occur in. These roles can have a more specific

meaning such as buyer in the frame Commerce-buy, but they can also refer to a more

general notions such as agent in the frame Accomplishment (see Section 2.2.2). VerbNet

uses a set of 23 roles with general meaning that are interpreted in the same way with

all verbs. PropBank uses only 6 role labels, but their interpretation varies depending

on the context (see Section 2.2.2). Interestingly, all the three resources adopt atomic

notions and relatively arbitrary role sets, despite the arguments for decomposing the

predicate-argument structure, put forward in the linguistic literature (see Section 2.1.2).

PropBank labels are based on Dowty (1991)’s notions of proto-roles, but the properties

68


which define to what degree a role belongs to one of the two types (see Section 2.1.3)

are not annotated separately.

A number of experiments have been conducted to investigate how the differences between

the PropBank and the VerbNet annotation schemes influence the systems for automatic

role labelling. The task of learning the VerbNet labels can be expected to be more

difficult, since there are more different items to learn. On the other hand, the fact that

the labels are used in a consistent way with different verbs could make it easier because

the labels should be better associated with the other features used by the systems.

Loper et al. (2007) show that the system trained on the VerbNet labels predicts better

the label for new instances than the system trained on the PropBank labels, especially

if the new instances occur in texts of a different genre. However, this finding only holds

if the performance is compared for the arg1 and arg2 labels in PropBank vs. the

sets of VerbNet labels that correspond to them respectively. The VerbNet labels were

grouped in more general labels for this experiment, 6 labels corresponding to arg1 and

5 corresponding to arg2. If the overall performance is compared, the PropBank labels

are better predicted, which is also confirmed by the findings of Zapirain et al. (2008).

Merlo and van der Plas (2009) compare different quantitative aspects of the two anno-

tation schemes and propose the ways in which the resources can be combined. They

first reconsider the evaluation of the performances of the systems for automatic seman-

tic role labelling. They point out that an uninformed system that predicts only one

role, the most frequent one, for every case would be correct in 51% cases if it learned

the PropBank roles, and only in 33% cases if it learned the VerbNet roles, due to the

different distributions of the instances of the roles in the corpus. They neutralise this

bias by calculating and comparing the reduction in error rate for the two annotations.

According to this measure, the overall performance is better for the VerbNet labels,

but it is more degraded in the cases where the verb is not known (not observed in the

training data) compared to the cases where it is known, due to the stronger correlation

between the verb and its role set. Thus, they argue that the VerbNet labels should be

used if the verb is known, and the PropBank labels if it is new.

Looking at the joint distribution of the labels in the corpus, Merlo and van der Plas

(2009) note different relations for roles with different frequencies. The frequent labels in

69


PropBank can be seen as generalizations of the frequent labels in VerbNet. For example

agent and experiencer are most frequently labelled as arg0, while theme, topic,

and patent are most frequently labelled as arg1, which means that the PropBank

labels group together the similar VerbNet labels. The PropBank labels of low frequency

are more specific and more variable, due to the fact that they depend on the context,

and the VerbNet labels are more stable. Thus for interpretation for a particular instance

of a PropBank label a VerbNet label could be useful.

The comparisons of the sets of labels used in PropBank and VerbNet annotation schemes

indicate that they can be seen as complementary sources of information about semantic

roles. However, other aspects of combining the resources are still to be explored. The

described comparisons are performed only for the lexical units which are included both

in PropBank and VerbNet. Since these two resources contain different lexical items, their

combination might be used for increasing the coverage. Also, the potential advantages of

using the other data provided by the resources (e.g. the hierarchies defined in FrameNet

and VerbNet) are still to be examined.

2.3. Automatic approaches to the predicate-argument

structure

The relational meaning of verbs, represented as the predicate-argument structure, is not

only interesting from a theoretical point of view. As a relatively simple representation of

the meaning of a clause, which can also be related with some observable indicators (see

Section 2.2.1), this representation has attracted considerable attention in the domain of

automatic analysis of the structure of natural language. An automatic analysis of the

predicate-argument structure can be useful for improving human-machine interface so

that computers can be exploited in searching for information in texts and databases,

automatic translation, automatic booking, and others. For example, an automatic rail-

way information system could use such an analysis to “understand” that from Geneva

denotes the starting point and to Montreux the destination of the request in (2.43).

(2.43) What is the shortest connection from Geneva to Montreux?

70

2.3. Automatic approaches to the predicate-argument structure

Automatic analysis of the predicate-argument structure relies on the observations of

the instances of verbs in large samples of texts, electronic corpora. The observable

characteristics of the instances of verbs are formulated as features which are then used

to train machine learning algorithms. Most of the algorithms used in computational

approaches to the predicate-argument structure are not tailored specifically for natural

language processing, but they are general algorithms which can be applied to a wider

range of machine learning tasks. Nevertheless, with well chosen feature representation,

statistic modelling of data, and an appropriate architecture, systems generally manage

to capture semantically relevant aspects of the uses of verbs showing high agreement

with human judgments.

In this dissertation, we regard computational approaches to the predicate-argument

structure of verbs as a suitable experimental framework for testing our theoretical

hypotheses. To achieve a good performance in automatic analysis of the predicate-

argument structure, it is necessary to capture generalisations about the relationship

between the meaning of verbs, which people interpret intuitively, and the distribution

of different observable characteristics of verb uses in language corpora. Since the same

relationship is modelled in our work, we study and apply methods used in computational

approaches.

As opposed to theoretical approaches, computational approaches put the accent on pre-

dictions rather than on the generalisations themselves. In theoretical accounts of lin-

guistic phenomena, generalisations are usually stated explicitly in the form of grammar

rules. In computational approaches, generalisations are often formulated in terms of

statistical models expressing explicitly relationships between structural elements, but

not necessarily in the form of grammar rules. Predictions which follow from the gener-

alisations can be explicitly formulated in theoretical accounts, but this is not necessary.

Contrary to this, predictions are precisely formulated in computational approaches and

tested experimentally.

Another important difference between theoretical and computational approaches is in

the theoretical context which is assumed for each particular problem. While theoreti-

cal accounts treat particular problems in relation to a more general theory of linguistic

structure, computational approaches are focused on specific tasks treating them as inde-

71


pendent of other tasks. The task orientation in computational approaches allows specific

definitions of predictions and measuring the performance, but theoretical relevance of

discovered generalisations is often not straightforward.

In this section, we discuss the work in the natural language processing framework which

involves analysing the predicate-argument structure of verbs. We concentrate especially

on two tasks which deal with the relationship between the meaning of verbs and the

structure of the clauses: semantic role labelling and verb classification. We briefly de-

scribe other related tasks which are less directly concerned with the relationship studied

in our own experiments.

2.3.1. Early analyses

Early work on automatic analysis of the syntactic realisations of verbs’ semantic argu-

ments centred around automatic development of lexical resources which would be used

for syntactic parsing and text generation. The work was based on the assumption that

the number of potential syntactic analyses can be significantly reduced if the information

about the verb’s subcategorisation frame is known (Manning 1993; Brent 1993; Briscoe

and Carroll 1997). It was soon understood that the notion of verb subcategorisation

alone does not capture the relevant lexical information. Due to the alternations of ar-

guments, many verbs are systematically used with multiple subcateogrisation frames.

The subsequent work brought some proposals for automatic identification of argument

alternations of verbs (McCarthy and Korhonen 1998; Lapata 1999). These proposals

still concerned mostly syntactic issues.

Early approaches, as well as the majority of the subsequent work on lexical and syn-

tactic properties of verbs, do not target the nature of the relationship between a verbal

predicate and its semantic arguments. The tasks are defined in terms syntactic sub-

categorisation and selectional preferences. The aim of this research is to improve the

performance of automatic parsers by limiting the range of possible syntactic constituents

with which a verb can be combined (the taks of identifying the subcategorisation frames)

and the range of possible lexical items which can head these constituents (the task of

identifying selectional preferences).

72


Identifying the nature of the semantic relationships between verbs and their arguments

was established as a separate task, called semantic role labelling. In the following section,

we discuss in detail the methods used in this task.

2.3.2. Semantic role labelling

The work on automatic semantic role labelling was enabled by the creation of the re-

sources described in Section 2.2.2, which provided enough annotated examples to be used

for training and testing the systems. Since the first experiments (Gildea and Jurafsky

2002), semantic role labelling has received considerable attention, which has resulted in

a variety of proposed approaches and systems. Many of the systems have been devel-

oped and directly compared within shared tasks such as the CoNLL-2005 shared task

on semantic role labelling (Carreras and Marquez 2005) and the CoNLL-2009 shared

task on syntactic and semantic dependencies in multiple languages (Hajic et al. 2009).

Most of the numerous proposed solutions follow what can be considered the standard

approach.

Standard semantic role labelling

The most widely adopted view of the task of automatic semantic role labelling is the

supervised machine learning approach defined by Gildea and Jurafsky (2002). The term

supervised refers to the fact that a machine learning system is trained to recognise the

predicate-argument structure of a clause by first observing a range of examples where

the correct structure is explicitly annotated.

(2.44)

73


a)

S

VP

NP/theme

a cake

V

made

NP/agent

Mary

b)

S

VP

NP

a cake

V

made

NP

Mary

The annotation guides the system in selecting the appropriate indicators of the structure.

The program reads the training input (a simplified example of a training sentence is

shown in (2.44a)) and collects the information about the co-occurrence of the annotated

structure and other observable properties of the phrase (lexical, morphological, and

syntactic). The collected observations are transformed into structured knowledge and

generalisations are made by means of a statistical model. Once the model is built using

the training data, it is asked to predict the predicate-argument structure of new (test)

phrases (illustrated in (2.44b)), by observing their lexical, morphological and syntactic

properties.

In the standard approach, the task of predicting the predicate argument structure of

a sentence is divided into two sub-tasks: identifying the constituents that bear a se-

mantic role (distinguishing them from the constituents that do not) and identifying the

semantic role label for all the constituents that bear one. Both sub-tasks are defined

as a classification problem: each constituents is first classified as either bearing or not

bearing a semantic role. Each constituent bearing a role is then classified into one of the

predefined semantic role classes. All the constituents belonging to the same class bear

the same semantic role. The two classification steps constitute the core of the semantic

role labelling task, which is usually performed in a pipeline including some pre- and

post-processing.

The pre-processing part provides the information which is considered given. First, the

predicates which assign semantic roles to the constituents are identified prior to semantic

role labelling proper. They are usually identified as the main verbs which head clauses.

Second, the syntactic analysis of the sentence that is being analysed is considered given.

Both pieces of information are obtained by morphological and syntactic processing of

74


the sentence. The relatively good performances of current morphological and syntactic

analysers allow these analyses to be performed automatically.

In practice, most of the systems use resources in which predicates are manually annotated

(see Section 2.2.2). However, Merlo and Musillo (2008) show that this information can

also be automated with comparable results, exploiting the relevant syntactic phenomena

already encoded in the syntactic annotation. In deciding which arguments belongs to

which predicate in a sentence, two sets of conditions are informative. First, the mini-

mality conditions determine whether another verb intervenes between a constituent and

its potential verb predicate. Second, the locality constraints determine whether the con-

stituent is realised outside of a verb phrase, either as the subject of a sentence or an

extracted constituent.

The post-procesing part can include various operations depending on the particular

system. In most cases, this part includes optimising at the sentence level. This step is

needed to account for the fact that semantic roles which occur in one sentence are not

mutually independent. For example, if one role is assigned to one syntactic constituent,

it is unlikely that the same role is assigned to another constituent in the same sentence.

In the standard approach, all the constituents are first assigned a role independently,

then the assignments are reconsidered in the post-processing phase taking into account

the information about the other roles in the sentence.

We do not discuss in detail all the aspects of automatic semantic role labelling, but

we focus on two aspects which are most relevant to our own experiments: knowledge

representation using features and statistical modelling of the collected data.

Features. The grammatical properties of phrases which are relevant for semantic role

labelling are described in terms of features.11 Different systems may use different fea-

tures, depending on the approach, but, as noted by Carreras and Marquez (2005) and

Palmer et al. (2010), a core set of features is used in almost all approaches. Those are

mainly the features defined already by Gildea and Jurafsky (2002):

11Note that the term feature is used in a different way in computational and in theoretical linguistics.Features in theoretical linguistics are more or less formal properties of lexical units which indicatewith what other lexical units they can be combined in syntactic derivations. In computationallinguistics, a feature can be any fact about a particular use of some word of a phrase.

75


• Phrase type — reflects the fact that some semantic roles tend to be realised by

one and others by another type of phrase. For example, the role goal tends to be

realised by noun phrases, and the role place is realised by prepositional phrases.

In training, the phrase type of each constituent annotated as a realisation of a

semantic argument of a verb is recorded. In the toy example given in (2.44), both

roles would be assigned the same value for this feature: NP.

• Governing category — defines the grammatical function of the constituent that

realises a particular semantic role. This feature captures the fact that some se-

mantic roles are realised as the subject in a sentence, and others are realised as the

direct object. The feature is defined so that it can only have two possible values:

S and VP. If a constituent bearing a semantic role is governed by the node S in a

syntactic tree, it is the subject of the sentence (Mary in (2.44)); if it is governed

by the node VP, it means that it belongs to the verb phrase, which is the position

of the object (a cake in (2.44)). The difference between the direct and the indirect

object is not made.

• Parse tree path — defines the path in the syntactic tree which connects a given

semantic role to its corresponding predicate. The value of this feature is the

sequence of nodes that form the path, starting with the verb node and ending

with the phrase that realises the role. The direction of moving from one node to

another is marked with arrows. For example, the value of the feature for the agent

role in the example (2.44) relating it to the verb made would be: V↑VP↑S↓NP; the

value of this feature for the theme role would be: V↑VP↓NP. The whole string is

regarded as an atomic value. Possible values for this feature are numerous. Gildea

and Jurafsky (2002) count 2 978 different values in their training data.

• Position — defines the position of the constituent bearing a semantic role rela-

tive to its corresponding predicate, whether the constituent occurs before or after

the predicate. This is another way to describe the grammatical function of the

constituent, since subjects tend to occur before and objects after the verb.

• Voice — marking whether the verb is used as passive or active. This feature is

needed to capture the systematic alternation of the relation between the gram-

matical function and semantic role of a constituent. While agent is the subject

76


and theme is the object in typical realisations, the reverse is true if the passive

transformation takes place.

• Head word — describes the relation between the lexical content of a constituent

and the semantic role that it bears. The value of this feature is the lexical item

that heads the constituent. For example, a constituent which is headed by Mary is

more likely to be an agent than a theme, while the constituent headed by cake

is more likely to be a theme, as it is the case in (2.44).

The overview of the features shows that the systems for automatic identification of the

roles of verbs’ arguments largely rely on the syntactic analysis and the relation between

the type of a semantic role and the syntactic form. Three of the listed features, path,

government, and position are different indicators of the grammatical function of the

constituents. Gildea and Jurafsky (2002) compare performances of the system using

only one or two features at a time to the performance using the whole set. They find

that using both the position and either of the other two feature is redundant. On the

other hand, including any of these features is necessary.

More recent systems use more features. In addition to the government feature, for

instance, the information about the siblings of the constituent in the tree is collected.

Also, information about the subcategorization frame or syntactic pattern of the verb is

often used (Carreras and Marquez 2005).

Selecting a particular set of features to represent the relevant knowledge about the

predicate-argument structure does not rely on any particular theoretical framework or

study. The choice of features tends to be arbitrary with little research on its linguistic

background. An exception to this is the study of Xue and Palmer (2004), which shows

that the feature set which should be used for argument identification is not the same as

the set which should be used for assigning the labels.

Modelling. When predicting the correct semantic role for a string of words (usually

representing a constituent of a sentence) the system observes the values of the defined

features in the test data and calculates the probability that each of the possible roles oc-

curs in the given conditions. The role that is most likely to occur in the given conditions

is assigned to the constituent.

77


The probability that is calculated for each possible role is formulated in the following

way (Gildea and Jurafsky 2002):

P (r|h, pt, gov, position, voice, t) (2.45)

The knowledge about the current instance which is being classified consists of the values

of the features listed on the right-hand side of the bar. The formula in (2.45) reads in the

following way: What is the probability that a particular constituent bears a particular

semantic role r knowing that the head of the constituent is h, the path between the

constituent and the predicate is pt, the category governing the constituent is gov, the

position of the constituent relative to the predicate is position, the voice of the verb

predicate is voice, and the verb predicate is t?

To choose the best role for a particular set of feature values, the probability of each

role in the given context needs to be assessed. One could assume that the role which

occurs most frequently with a given combination of values of the features in the training

data is the role that is most likely to occur with the same features in the test data

too. In this case, the probability could be calculated as the relative frequency of the

observations: the number of times the role occurs with the combination of features out

of all the occurrences of the combination of features in question:

P (r|h, pt, gov, position, voice, t) =#(r, h, pt, gov, position, voice, t)

#(h, pt, gov, position, voice, t)(2.46)

The problem with this approach is that some features can have many different values

(e.g. the value of the feature head word can be any word in the language), which results

in a large number of possible combinations of the values. Many of the combinations will

not occur in the training data, even if large-scale resources are available. Thus, the set

of features has to be divided into subsets that occur enough times in the training data.

The values for each subset are then considered for each possible semantic role and the

decision on the most probable role is made by combining the information. Gildea and

Jurafsky (2002) divide the set of features into 8 subsets:

78


P (r|t), P (r|pt, position, voice, t),P (r|pt, t), P (r|h),

P (r|pt, gov, t), P (r|h, t),P (r|pt, position, voice), P (r|h, pt, t).

They explore several methods of combining the information based on the subsets achiev-

ing the best results by combining linear interpolation with the back-off method. Linear

interpolation provides the average value of the probabilities based on the subsets of

features. It is calculated in the following way:

P (r|constituent) =λ1P (r|t) + λ2P (r|pt, t)+

λ3P (r|pt, gov, t) + λ4P (r|pt, position, voice)+

λ5P (r|pt, position, voice, t) + λ6P (r|h)+

λ7P (r|h, t) + λ8P (r|h, pt, t) (2.47)

where λi represents interpolation weight of each of the probabilities and Σiλi = 1.

It can be noted that not all subsets include the same number of features. By including

more features, the subset (pt, position, voice, t), for instance, defines more specific con-

ditions than the subset (t). The back-off method enables combining the more specific

features, that provide more information, when they are available and turning to the

more general features only if the specific features are not available. The values for the

most specific subsets ((pt, position, voice, t), (pt, gov, t), (h, pt, t)) are considered first. If

the probability cannot be estimated for any of them, it is replaced by its corresponding

less specific subset. For example, (pt, position, voice, t) is replaced by (pt, position, t),

(pt, gov, t) by (pt, t), (h, pt, t) by (h, t) and so on.

Different systems apply different machine learning methods to estimate the probabilities.

A range of different methods, including those based on maximum entropy, support vector

machines, decision tree learning and others, have been applied in more recent systems

(Carreras and Marquez 2005) .

79


The described classification applies to the task of assigning a semantic role to a con-

stituent which is known to bear one. The same methods can be used for the first step

in semantic role labelling, that is to identifying the constituents which bear semantic

roles. The estimated probability in this case is the probability that a constituent bears

any semantic role in given conditions, described by a reduced set of features. Gildea and

Jurafsky (2002) use the information on the head word (feature h), target word t and the

path between them.

Joint and unsupervised learning

There are two kinds of approaches which can be seen as not following the standard

pipeline framework. One line of the development explores the potential of joint mod-

elling. These statistical models and computational methods exploit the relationship

between the syntactic and the predicat-argument structure in a more systematic way.

Toutanova et al. (2005) shows that moving the account of the global outline of a sentence

from the post-processing phase to the core statistical model improves the classification

results. Henderson et al. (2008) propose a model for joint learning of both syntactic

and semantic labelling in a single model, moving the syntactic information from the

pre-processing phase to the core statistical model. The advantage of such approaches

compared to the standard approach is that the syntactic structure of a phrase is not

definitely assigned before the semantic structure, so that the semantic information can

be used for constructing a better syntactic representation, reducing error propagation

between the levels of analysis.

Recently, the attention has been focused on unsupervised learning, where the information

about correct semantic role labels (assigned by human annotators) is not available for

training. The advantage of unsupervised approaches (Lang and Lapata 2011; Titov and

Klementiev 2012; Garg and Henderson 2012) compared to the standard approach is

that they do not require manually annotated training data, which are costly and hard to

develop (see Section 2.2.2 for more detail). Unsupervised learning exploits the overlap

between syntactic representation and the predicate-argument structure. The models

cluster the instances of syntactic constituents described in terms of features (similar to

the features used in the standard approaches). The constituents which are similar in

80


terms of their feature representations are grouped together. The models include a hidden

layer representing semantic roles which potentially underlie the observed distribution of

the constituents.

2.3.3. Automatic verb classification

The task of automatic verb classification addresses not only the predicate-argument

structure of verbs, but also the semantic classification of verbs. An in-depth analysis

of the relationship between the lexical semantics of verbs and the distribution of their

uses in a corpus is performed by Merlo and Stevenson (2001). The study addresses the

fine distinctions between three classes of verbs which all include verbs that alternate

between transitive and intransitive use. The classes in question are manner of motion

verbs (2.48), which alternate only in a limited number of languages, change of state

verbs (2.49), alternating across languages, and performance/creation verbs (2.50).

(2.48) a. The horse raced past the barn.

b. The jockey raced the horse past the barn.

(2.49) a. The butter melted in the pan.

b. The cook melted the butter in the pan.

(2.50) a. The boy played.

b. The boy played soccer.

Although the surface realisations of phrases formed with these verbs are the same (they

all appear both in transitive and intransitive uses), the underlying semantic analysis of

the predicate-argument structure is different in each class. The subject of intransitive

realisation is agentive (animate, volitional) in (2.48a) and (2.50a), while it is not in

(2.49a). On the other hand, the transitive realisation contains one agentive and one

non-agentive role in (2.49b) and (2.50b), while the realisation in (2.48b) contains two

agentive arguments. Correct classification of verbs into one of the three classes defines,

thus, the correct analysis of the semantic relations that they express.

81


Merlo and Stevenson (2001) base their approach to automatic classification of verbs on

the theoretical notion of linguistic markedness. The main idea of the theory of marked-

ness is that linguistic marking occurs in elements which are unusual, unexpected, while

the common, expected elements are unmarked. To use a common simple example, the

plural of nouns in English is marked with an ending (’-s’) because it is more uncommon

than singular which is unmarked. Linguistic markedness has direct consequences on

the frequency of use: it has been shown that marked unites are rarer than unmarked

unites.

Applied to the choice between the intransitive and transitive use of the verbs addressed

by Merlo and Stevenson (2001), the theory of linguistic markedness results in certain

expectations about the distribution of the uses. It can be expected, for example, that

the uses such as (2.48a) are unmarked, which means more frequent, while the uses

such (2.48b) are marked, which means less frequent. In the case of verbs represented in

(2.50) the expected pattern is reversed: the intransitive use (2.50a) is marked here, while

the transitive use (2.50b) is unmarked. For the verbs illustrated in (2.49) none of the

uses is marked, which means that roughly equal number of transitive and intransitive

realisations is expected.

Features. In the classification task, the uses of verbs are described in terms of features

which are based on a combination of the markedness analysis with an analysis of semantic

properties of the arguments of the verbs. Three main features are defined:

• Transitivity — captures the fact that transitive use is not equally common for

all the verbs. It is very uncommon for manner of motion verbs (2.48b), much

more common for change of state verbs (2.49b), and, finally, very common for

performance/creation verbs (2.50b). This means that manner of motion verbs are

expected to have consistently a low value for this feature, change of state verbs

middle, and performance/creation verbs high.

• Causativity — represents the fact that, in the causative alternation, the same

lexical items can occur both as subjects and as objects of the same verb. This

can be expected for arguments such as butter in (2.49) and horse in (2.48). This

feature is operationally defined as the rate of overlap between lexical units found

as the head of the subject of the intransitive uses and those found as the head

82


of the object in the transitive uses of the same verb. The quantity is expected

to distinguish between the two classes illustrated in (2.48) and (2.49) on one side

and the class illustrated in (2.50) on the other side, because the alternation in the

latter class is not causative (the object of the transitive use does not appear as the

subject of the intransitive use, it is simply left out).

• Animacy — is used to distinguish between the verbs that tend to have animate

subjects (manner of motion verbs (2.48) and performance verbs (2.50)) and those

that do not (change of state verbs (2.49)). It is operationally defined as the rate

of personal pronouns that appear as the subjects of verbs.

Additional features are also used (the information about the morphological form of the

verb) but they are not as theoretically prominent as the main three features.

Classification. The experiments in classification are performed on 60 verbs (20 per

class) listed as belonging to the relevant verb classes by Levin (1993). Each verb is

described as a vector of feature values, where the values are calculated automatically

from corpus data, as shown for the verb form opened in (2.51).

(2.51) a)verb trans pass vbn caus anim class-code

opened .69 .09 .21 .16 .36 unacc

b)verb trans pass vbn caus anim class-code

opened .69 .09 .21 .16 .36 ?

The co-occurrence of the feature values with a particular class is observed in the training

data and registered. The training input is illustrated in (2.51a). The first six positions

in the vector represent the values of the defined features extracted from instances of the

verb in a corpus. The the last position in the vector is the class that should be assigned

to the verb. The code unacc refers to the term unaccusative verbs, which is often used

to refer to the change-of-state verbs. In predicting the class that is assigned to a verb

in the test input (illustrated in (2.51b)), the probability of each class being associated

with the observed vector of feature values is assessed. The algorithm used for calculating

the most probable class is a supervised learning algorithm, the decision tree, which is

described in more detail in Chapter 3.

83


The results of the study show that the classifier performs best if all the features are

used. They also show that the discriminative value of the features differs when they are

used separately and when they are used together, which means that information about

the use of verbs that they encode is partially overlapping. Subsequent studies develop

in different directions. While Merlo et al. (2002) explore using cross-linguistic informa-

tion as a kind of additional general supervision in the classification task, most of the

remaining work concerns two interrelated lines of research: unsupervised classification

and generalisation.

Lapata and Brew (2004) propose a statistical model of verb class ambiguity for unsuper-

vised learning the classification preferences of verbs which can be assigned to multiple

classes. The model does not use a predefined set of linguistically motivated features as in

the approach of Merlo and Stevenson (2001), but it takes into account the distribution

of a wide range of verbs (those listed by Levin (1993)) and their syntactic realisations

in combination with the distribution of classes. The resulting preferences are then used

to improve verb senses disambiguation.

Several studies deal with the required feature set (Stevenson and Joanis 2003; Joanis and

Stevenson 2003; Joanis et al. 2008; Schulte im Walde 2003; Schulte im Walde 2006;

Li and Brew 2008), especially in the unspervised and partially supervised setting. This

work suggests that the set of features which is useful for verb classification is not specific

to this task. Schulte im Walde (2003) argues that no generally useful features can be

identified, but that the usefulness of a feature depends on the idiosyncratic properties of

verb classes. Baroni and Lenci (2010) explore further potential generalisations in lexical

acquisition from corpora, proposing a framework for constructing a general memory

of automatically acquired lexical knowledge about verbs. This knowledge can be used

directly for different classifications required by different applications. Schulte im Walde

et al. (2008), Sun and Korhonen (2009) explore further the effects of incorporating the

information about lexical preferences of verbs into verb classification, which had proved

to be less helpful than expected in earlier experiments (Schulte im Walde 2003).

84

2.4. Summary

2.4. Summary

We have shown in this chapter how the view of the lexical representation of verbs has

evolved in linguistic theory and how it was followed in computational linguistics. Three

turning points in theoretical approaches to the meaning of verbs can be identified. First,

the relational meaning of verbs is separated from the other, idiosyncratic semantic com-

ponents. The relational meaning, called the predicate-argument structure, is then further

analysed. There are two main approaches to the decomposition of the predicate argu-

ment structure: decomposition of the arguments into sets of features and decomposition

of the verbal predicates into sets of predicates. In the latter approach, an attempt is

made to derive the decomposition from more general semantic templates (such as causal

or temporal).

The predicate-argument structure has recently been recognised in computational lin-

guistics as a level of linguistic representation that is suitable and useful for automatic

analysis. The view of the predicate argument structure underlying the computational ap-

proaches, however, does not follow the developments in linguistic theory. The overview of

the knowledge representation in the resources used for training automatic systems shows

that the predicate-argument structure which is annotated and automatically learnt is

still based on the atomic view of the predicates and arguments, despite the fact that this

view is shown to be theoretically inadequate in the linguistic literature. The feature-

based knowledge representation used in the statistical models is also not closely related

to the notions discussed in the linguistic literature. However, the work on automatic se-

mantic role labelling and verb classification shows the potential of using the information

about verb instances in corpora for recognising fine components of verbal meaning.

In this dissertation, we use the methods developed in the approaches to automatic

acquisition of the meaning of verbs to learn automatically the components of the lexical

representation which are relevant to the current discussion in linguistic theory. The

components of verbs’ meaning which we identify on the basis of the distributions of their

syntactic realisations observed in a corpus are defined in terms of causal and temporal

decomposition of events described by the verbs.

Studying the uses of verbs in parallel corpora is the main novelty of this work. By

85


taking this approach, we make a step further with respect to both existing computational

approaches and the standard linguistic methodology. Previous corpus-based explorations

of the meaning of verb are generally monolingual and they do not address the patterns in

the cross linguistic variation. On the other hand, cross-linguistic data are crucial for the

standard methodology of linguistic research. However, the standard approaches usually

involve just a few instances of the studied phenomena which are discussed in depth. In

contrast to this, the approach which we propose allows studying cross-linguistic variation

in a more systematic way, taking into consideration large data sets. The details of our

approach based on parallel corpora are discussed in the following chapter.

86

3. Using parallel corpora for linguistic

research — rationale and

methodology

A parallel corpus is a collection of translations between two (or more) languages, where

each sentence in one language is aligned with the corresponding sentence in the other

language. The work on constructing numerous parallel corpora of different languages was

primarily motivated by the developments in statistical machine translation in the early

nineties. With the emergence of systems able to learn to translate from one language

to another by observing a set of examples of translated sentences, resources for training

and evaluating such systems started growing rapidly. Current versions of some popular

sentence-aligned parallel corpora, such as Europarl (Koehn 2005) or OPUS (Tiedemann

2009), contain tens of languages, with some languages being represented with millions of

sentences. These resources are still used mostly for experiments in machine translation,

but potential other uses are increasingly proposed and explored.

In this dissertation, parallel corpora are used for investigating lexical representation of

verbs. To address theoretical questions concerning the meaning of verbs, we design a

novel methodology which combines methods originating in several disciplines. We for-

mulate our research hypotheses on the basis of theoretical discussions and arguments

put forward in the linguistic literature. We then collect a large number of empirical

observations relevant to the hypotheses from parallel corpora using state-of-the-art nat-

ural language processing. We perform different statistical analyses of the collected data.

Some interesting insights are obtained by a simple descriptive analysis where a summary

of a large number of observations reveals significant patterns in the use of verbs. In some

87

3. Using parallel corpora for linguistic research — rationale and methodology

cases, we use standard statistical tests to determine whether some identified tendencies

are statistically and scientifically significant. When a more complex analysis is required,

we design statistical models, which are intended to explain the observations with a set of

generalisations. To test the predictive performance of the models, we employ standard

machine learning methods which are commonly used in natural language processing, but

not in theoretical linguistics. We train the models on a large set of examples of verbs’

use extracted from parallel corpora using machine learning algorithms. We then test and

evaluate the predictions made by the models on an independent set of test examples.

The methods are described in more detail in the remaining of this chapter.

The chapter consists of two major parts. In the first part (Section 3.1 and Section 3.2) we

discuss several methodological issues related to parallel corpora. We start by presenting

our arguments for using parallel corpora for linguistic research, showing that our method

based on parallel corpora can be regarded as an extension of the standard theoretical

approach to cross-linguistic variation (3.1.1). We then discuss potential methodological

problems posed by translation effects, which can influence the representativeness of par-

allel corpora (3.1.2). In Section 3.2, we first discus in detail automatic word alignment,

which is crucial for automatic extraction of data from parallel corpora (3.2.1). We then

give a brief overview of how parallel corpora have been used for research in natural

language processing as in illustration of the potential of parallel corpora as sources of

linguistic data (3.2.2). In the second part (Section 3.3 and Section 3.4), we present

the technical details of the methodology which we apply to analyse the data extracted

from parallel corpora. Section 3.3 contains an introduction to statistical inferences and

modelling. In Section 3.4 we lay out machine learning approaches to training statisti-

cal models, providing more details about the learning algorithms which are used in the

experiments in this dissertation.

3.1. Cross-linguistic variation and parallel corpora

In theoretical linguistics, cross-linguistic variation has always been studied as a means of

discovering elements of linguistic structure which are invariably present in all languages,

the atoms of language as metaphorically put by Baker (2001). Linguistic analysis almost

88


inevitably involves parallel sentences such as the pair Gungbe-English in (3.1), taken

from Aboh (2009), or the pair Ewe-English in (3.2) by Collins (1997).

(3.1) AsıbaAsiba

dacook/prepare/make

lεsırice

du.eat

Asiba cooked the rice eat(i.e. she ate the rice).

(3.2) KofiKofi

tsotake

ati-εstick-def

fohit

YaoYao

(yi).P

Kofi took the stick and hit Yao with it.

Such parallel sentences are usually constructed on the basis of native-speaker competence

to illustrate apparently different realisations of a particular construction in different

languages ((3.1) and (3.2) are examples of complex predicates) and to identify the level

of representation at which the languages do not differ. For the reason of simplification,

we do not show the full analysis of the examples (3.1) and (3.2), but they illustrate

a situation where the same kind of complex predicate-argument structures are realised

with two separate clauses in English, and with a single clause in Gungbe and Ewe. In

this case, the predications expressed in the sentences are invariable across languages,

while the structural level at which they are realised is varied.

Parallel sentences cited and analysed in the linguistic literature usually represent the

most typical realisations, abstracting away from potential variation in realisations in

both languages. The corpus of analysed cases rarely exceeds several examples for each

construction studied.

3.1.1. Instance-level microvariation

In this dissertation, parallel realisations of particular constructions are studied on a

much larger scale taking into consideration the potential variation in realisations. We see

parallel corpora as samples of sentences naturally produced in two (or more) languages,

from which we extract all instances of the studied constructions, and not just typical

uses, relying on statistical methods in addressing the variation. This approach allows

us to observe many different realisations of constructions that actually occur in texts

89


and to address the non-canonical uses as well as the canonical realisations. Since we

work with actual translations, the cross-linguistically equivalent expressions are directly

observed. We do not have to rely on our intuition about which construction in one

language corresponds to which construction in the other language. We simply observe the

realisations in the aligned sentences and then summarise (or classify) the observations. In

this way, we can identify grammatically relevant tendencies which cannot be observed

using standard approaches. For example, passive constructions are available both in

English and German and they can be seen as equivalent forms. However, verbs in one

of the two languages may show a tendency to be realised in passive forms in the same

context where intransitive realisations are preferred by the other language. Such an

asymmetry might prove to be grammatically relevant.

Studying instances of verbs in a parallel corpus makes it possible to control for any

pragmatical and contextual factors that may be involved in a particular realisation

of a verb, allowing us to isolate structural factors which underlie the variation in the

realisations. Since translation is supposed to express the same meaning in the same

context, we can assume that the same factors that influence a particular realisation

of a verb in a clause in one language influence the realisation of its translation in the

corresponding clause in another language. Any potential differences in the form of the

two parallel clauses should be explained by the lexical properties of the verbs or by

structural differences between languages.

Studying many different instances of verbs in parallel corpora fits well with some re-

cent general trends in theoretical linguistics. In the current methodology of linguistic

research, small lexical variation between similar languages has been given an important

place. As discussed in several places in a collection of articles consecrated to the theoret-

ical aspects of cross-linguistic variation (Biberauer 2008), a distinction is made between

macro-parameters and micro-parameters.

In making this distinction, the term macro-parameters is used for those parameters

of variation which are traditionally studied, mostly in the framework of the theory

of principles and parameters (Chomsky 1995). Such a parameter is, for example, the

famous pro-drop parameter, which divides languages into two major groups: those where

expressing the subject of a sentence is obligatory, such as English and French, and those

90


where the subject can be omitted when it is expressed by a pronoun (hence the term

pro-drop), such as Italian. The term macro-parameter does not only refer to the fact

that these parameters concern all (or almost all) languages, but also to the fact that

they concern large structural chunks. Presence vs. absence of the subject is the kind of

variation that affects the basic layout of sentences, causing substantial differences in the

structure of sentences across languages.

As opposed to macro-parameters, micro-parameters concern the variation which is lim-

ited to smaller portions of sentences. They affect the structure of small phrases and,

especially, the choice of lexical items. They also apply to a smaller number of languages.

Micro-parameters are typically studied when structures are compared between closely

related languages, which have the same setting of macro-parameters. An example of a

micro-parametric category is the difference between the French quantifier beacoup and

its apparently corresponding English quantifier many. In an influential study, Kayne

(2005) shows that the two lexical items have different representations, although they are

considered equivalent. Kayne’s study is set within the programme of isolating minimal

abstract units of language structure by identifying minimal structural divergence in two

similar languages or even dialects of the same language.

We see parallel corpora as a suited resource for studying micro-variation. Numerous

examples of uses of lexical items can be extracted from parallel corpora and studied in a

systematic way. Applying automatic methods for extraction allows us to analyse not only

many instances of lexical items, but also many items. While theoretical investigations

are usually limited to just several items which are analysed at the type level, our studies

include thousands of instances of hundreds of verbs, which provides a strong empirical

basis for new theoretical insights. We underline that this advantage applies only to

those phenomena which are frequent enough so that a sample of instances can be found

in corpora. Lexical items with grammatically relevant properties, like the quantifiers

studied by Kayne (2005) and the verbs studied in this dissertation, represent exactly

that kind of linguistic phenomena.

Although cross-linguistic variation is one of the crucial issues in linguistic theory, parallel

corpora are rarely used in linguistic research outside natural language processing. The

importance of parallel corpora for linguistic research has been recognised mostly by the

91


researchers in the domain of language typology. A collection of papers edited by Cysouw

and Walchli (2007) brings several case studies demonstrating the kind of language facts

that can be extracted from parallel corpora. A broader study is performed by von

Waldenfels (2012) who uses a parallel corpus of eleven Slavic languages to study the

variation in the use of imperative forms. The patterns that are found in the corpus

data are shown to correspond to the traditional genetic and areal classification of Slavic

languages.

Linguistic investigations of parallel corpora are not only rare, but they are also lit-

tle automated. Data collection and, especially, analyses are performed almost entirely

manually, which means that the number of observations which can be analysed is rather

small compared to the available information in the resources. In contrast to this, the

methodology proposed in this dissertation is entirely automatic, drawing heavily on the

approaches in natural language processing.

3.1.2. Translators’ choice vs. structural variation

One important limitation of using parallel corpora for linguistic research is the fact

that, despite controlling for context and discourse factors, translations might still include

variation which is not necessarily caused by linguistic divergences. Consider, for example,

the English sentence in (3.3a) and its French translation in (3.3b).

(3.3) a. I hope that the President of the Commission [...] tells us what he intends to

do.

b. J’espereI hope

quethat

lethe

presidentpresident

deof

lathe

Commissioncommission

[...] nousus

ferawill make

partpart

deof

seshis

intentions.intentions

Even if the English sentence in (3.3a) can be translated into French keeping the parallel

structure, it is not. As a result, the phrases tells us what he intends to do and nous fera

part de ses intentions (will make us part of his intentions) cannot be seen as structural

92


counterparts, although the two languages can express the content in question in a struc-

turally parallel way. There is a verb in French (communiquer) that corresponds to the

English verb tell, taking the same types of the complements as the English verb. How-

ever, at the instance level, these two sentence are not parallel. The factors that influence

the translations at the instance level are numerous, including discourse factors, broader

cultural context, translators attitudes, and other factors. An interesting question to ask,

then, is to what degree the existing translations actually show the structural divergence

between languages.

In an experimental study on a sample of 1 000 sentences containing potentially parallel

frames in the sense of FrameNet (see Section 2.2.2 in Chapter 2), extracted from the

Europarl corpus and manually annotated, Pado (2007) finds that 72% of English frames

that could have a parallel frame in German were realized as parallel instances. The ratio

is 65% for the pair English-French. However, once the frames are parallel, the parallelism

between the roles (frame elements in FrameNet) within the frames is assessed as “almost

perfect”.

We address this limitation by extracting only the most parallel sentences. We use the

information obtained by automatic alignment of words in parallel sentences and auto-

matic linguistic analysis of the sentences on both sides of a parallel corpus (described in

more detail in Section 3.2.1) to control the kind of constructions which are extracted.

We extract only the realisations which show certain levels of parallelism, minimising the

variation which is potentially irrelevant for linguistic studies.

Another potential problem for using parallel corpora for linguistic research are the known

translation effects. It has been argued that the language of translated texts differs

from the language of texts originally produced in some language in several respects.

Baroni and Bernardini (2006) have shown, for example, that, given a choice between an

expressions which is similar to the one in the source language and an expression which is

different, translators tend to chose the different expression. The result of this tendency is

more divergence in the translations than it is imposed by structural differences between

the languages. Also, different translators might have different strategies in choosing the

expressions.

93


This limitation is partially addressed by the strategy of maximising parallelism in ex-

tracting the instances of verbs. Another strategy that we apply to addresses this issue is

using large-scale data. It can be expected that the variation which represents noise for

a linguistic analysis is marginalised in a big sample of instances which includes transla-

tions produced by many different translators. Patterns observed in a big sample can be

assigned to linguistic factors. The reasoning behind this expectation is that translators’

choice of expression is still limited by linguistic factors: they can only choose between

options provided by structural elements available in a language.

3.2. Parallel corpora in natural language processing

Collecting large data samples, which is crucial for studying parallel corpora, necessarily

involves automatic processing of texts. The information which is crucial for collecting

the data on cross-linguistic realisations of verbs is word alignment. If we want to ex-

tract verbs that are translations if each other in parallel sentences, the sentences need

to be word-aligned, so that, for each word in the sentence in one language, we know its

corresponding word in the sentence in the other language. Given that collecting large

samples implies extracting verb instances from hundreds of thousands of parallel sen-

tences, the required information can only be obtained automatically. In this section, we

discuss methods for automatic alignment of words in parallel corpora which have been

developed in the context of statistical machine translation.

3.2.1. Automatic word alignment

Word alignment establishes links between individual words in each sentence and their

actual translations in the parallel sentence. Figure 3.1 illustrates such an alignment,

where the German pronoun ich is aligned with the English pronoun I, German verb

form mochte, with the English forms would like and so on. As the example in Figure 3.1

shows, correspondences between the words in sentences are often rather complex. The

range of existing alignment possibilities can be described with the following taxonomy:

94


Figure 3.1.: Word alignment in a parallel corpus

• One-to-one alignment is the simplest and the prototypical case, where corre-

sponding single words are identified, such as I � Ich or lesson � Lehre in Figure

3.1

• One-to-null alignment can be used to describe words which occur in one language

but no corresponding words can be identified in the other language. In the example

in Figure 3.1, such words are English There, is, to and German daß.

• One-to-many alignment holds between a single word in one language and multiple

words in the other language. Examples of this relationship in Figure 3.1 are mochte

� would like and daraus � from this.

• Many-to-many alignment is necessary when no single word in any of the aligned

sentences can be identified as an alignment unit. This is usually case in aligning

idioms. The sentences in Figure 3.1 do not contain such an example. To illustrate

this case, we use the example in (3.4) taken from Burchardt et al. (2009). The

phrase nehmen in Kauf aligns with English put up with, but they can only be

aligned in the many-to-many fashion because no subpart of neither expression can

be identified as an alignment unit.

(3.4) a. Die Glaubiger nehmen Nachteile in Kauf. (German)

95


b. The creditors put up with disadvantages. (English)

Note that identifying alignments between words and phrases is not always straightfor-

ward. Although it is clear that units smaller than a sentence can be aligned, it is not

always clear what kind of alignment holds and between which words exactly. As an

illustration, consider the word to in the English sentence in Figure 3.1. Its alignment

is subject to interpretation. It can be seen as not corresponding to any word in the

parallel German sentence (one-to-null alignment), which is how it is aligned in our ex-

ample. However, since to marks the infinitive form in English and the corresponding

German verb is in the infinitive form, the one-to-many alignment to learn � ziehen is

also correct.

The alignment between English learn and German ziehen illustrates an important dif-

ference between word alignments and lexical translations. The two verbs are clearly

aligned in the example in Figure 3.1, but they are not lexical translations of each other.

Outside of the given context, German ziehen would translate to English draw or pull,

while English learn would translate to German lernen.

For the purpose of automatic extraction from parallel corpora, word alignment is usually

represented as a set of ordered pairs, which is a subset of the Cartesian product of the

set of words of the sentence in one language and the set of words of the aligned sentence

in the other language (Brown et al. 1993). Technically, one language is considered the

source and the other the target language, although this distinction does not depend on

the true direction of translation in parallel corpora. With the words being represented

by their position in the sentence, the first member in each ordered pair is the position

of the target word (j in 3.5) and the second member is the position of the source word

that the target word is aligned with (i in 3.5).

A ⊆ {(j, i) : j = 1, ..., J ; i = 1, ..., I} (3.5)

The set A is generated by a a single-valued function which maps each word in the target

sentence to exactly one word in the source sentence. For example, taking English as the

target and German as the source language in Figure 3.1, the alignment between I and

ich can be represented as the ordered pair (6, 1). Alignment of would like with mochte,

96


is represented with two ordered pairs (7, 2) and (8, 2). To account for the fact that

some target language words cannot be aligned with any source language word, a special

empty word (“NULL”) is introduced in the source sentence. In this way, all the words

that have no translation (such as English There, is, to in Figure 3.1) can be aligned with

this word, satisfying the general condition which requires that they are aligned with one

word in the source sentence.

Note that the given formal definition only approximates the intuitive notion of word

alignment described above. The definition simplifies the listed alignment relations in

two ways. First, one-to-many alignments are possible only in one direction; one source

word can be aligned with multiple target words, but not the other way around. As

a consequence, switching the target-source assignment of a pair of sentences changes

the alignment. Second, the single-valued function definition excludes many-to-many

relations entirely. Despite these limitations, the described formalisation is widely used

because it expresses the main properties of word alignment in a way that is suitable for

implementing algorithms for its automatic extraction from parallel corpora.

Word alignment is usually computed from sentence alignment by means of the expectation-

maximisation algorithm. The algorithm considers all possible alignments of the words in

a pair of sentences (the number of possible word alignment pairs is length(source))(length(target))

and outputs the one which is most probable. The probability of alignments is assessed

at the level of a sentence. Individual words are aligned so that the alignment score of

the whole sentence is maximised. The algorithm starts by assigning a certain initial

probability to all possible alignments. The probabilities are then iteratively updated on

the basis of observations in a parallel corpus. If a pair of words is observed together

in other pairs of sentences, the probability of aligning the two words increases. The

algorithm is described in more detail in Section 3.4.2.

A commonly used program that provides automatic word alignment of parallel corpora,

GIZA++ (Och and Ney 2003)), which is also used in our experiments, assumes the

alignment definition described above. In addition to the described basic elements (in-

dividual word alignment and global sentence alignment), the system implements some

refinements, which improve its actual performance. We do not discuss these refinements

since they do not introduce major conceptual changes.

97


The experiments performed to evaluate this alignment method (Och and Ney 2003)

showed that, apart from setting the required parameters, the quality of alignment de-

pends on the language pair, as well as on the direction of alignment (e.g. the performance

is better for the direction English → German than the other way around). They also

showed that combining the alignments made in both directions has a very good effect

on the overall success rate.

3.2.2. Using automatic word alignment in natural language

processing

Since parallel corpora have been available to the research community they have inspired

research in natural language processing beyond machine translation. A number of pro-

posals have been put forward to exploit translations of words automatically extracted

from parallel corpora for improving performance on different natural language process-

ing tasks. The work on part-of-speech tagging (Snyder et al. 2008) shows that data

from another language can help in disambiguating word categories. For example, the

English word can is ambiguous between three readings: it can be a modal verb, a noun,

or a lexical verb. Each of the three categories is translated with a different word in

Serbian, for example: the corresponding modal is moci, the noun is konzerva, and the

lexical verb is konzervirati. Knowing what is Serbian translation of the English word in

a given sentence can help decide which category to assign to the word.

The work of van der Plas and Tiedemann (2006) shows that the data from parallel cor-

pora can improve automatic detection of synonyms. The main difficulty for monolingual

approaches is distinguishing synonyms from other lexical relations such as antonyms,

hyponyms, and hypernyms, which all occur in similar contexts. For example, a mono-

lingual system would propose as synonyms the words apple, fruit, and pear because

they all occur in similar contexts. However, the fact that the three words are consis-

tently translated with different words into another language indicates that they are not

synonyms.

The potential of the data from parallel corpora for reducing ambiguity at different levels

of natural language representation has been used to improve syntactic analysis (Kuhn

98


2004; Snyder et al. 2009; Zarrieß et al. 2010), the analysis of the predicate-argument

structure (Fung et al. 2007; Wu and Palmer 2011), as well as machine translation

(Collins et al. 2005; Cohn and Lapata 2007).

An interesting application of parallel corpora is transferring structural annotation (mor-

phological, syntactic, semantic) from one language to another. Developing resources

such as FrameNet or PropBank (see Chapter 2, Section 2.2.2), which have enabled

progress in automatic predicate-argument analysis, requires substantial investments in-

volving linguistic expertise, financial support, and technical infrastructure. This is why

such resources are only available for a small number of languages. Parallel corpora have

been seen as a means of automatic development of the resources for multiple languages.

The assumption behind the work on transferring annotation is that languages share

abstract structural representations and that whatever analysis applies to a sentence in

one language should be applied to its translation in another language. This assumption

is generally shared by theoretical linguists, as discussed in more detail in Section 3.1.

However, when tested on large corpora, the portability of a structural annotation is not

straightforward (Yarowsky et al. 2001; Hwa et al. 2002; Pado 2007; Burchardt et al.

2009; van der Plas et al. 2011). The work on automatic annotation transfer, although

primarily motivated by more practical goals, has provided some general insights con-

cerning the difference between the elements of the structure which are universal and

those which are language-specific.

The issue of parallelism vs. variation in the predicate-argument structure between En-

glish and Chinese is addressed by Fung et al. (2007), who study a sample of the Parallel

English-Chinese PropBank corpus containing over 1 000 manually annotated and man-

ually aligned semantic arguments of verbs (Palmer et al. 2005b). They find that the

roles do not match in 17.24% cases. English arg0 role (see Section 2.2.2 in Chapter

2) for more details), for instance, is mapped to Chinese arg1 77 times. Although the

sources of the mismatches are not discussed, the findings are interpreted as evidence

against the assumption that this level of linguistic representation is shared in the case

of English and Chinese.

The plausibility of a strong version of the assumption of structural parallelism is explored

by Hwa et al. (2002). It is formulated as the Direct Correspondence Assumption:

99


Given a pair of sentences E and F that are (literal) translations of each other

with syntactic structures TreeE and TreeF , if nodes xE and yE of TreeE

are aligned with nodes xF and yF of TreeF , respectively, and if syntactic

relationship R(xE, yE) holds in TreeE, then R(xF , yF ) holds in TreeF .

The evaluation of the annotation transferred from English to Chinese against a man-

ually annotated Chinese gold standard shows that syntactic relations are not directly

transferable in many cases. However, a limited set of regular transformations can be

applied to the result of direct projection to improve significantly the overall results. For

example, while English verb tense forms express verbal aspect at the same time (whether

the activity denoted by the verb is completed or not), Chinese forms are composed of

two words, one expressing the tense and the other the aspect. Projecting the annotation

from English, the relation between the aspect marker and the verb in Chinese cannot

be determined, since the aspect marker is either aligned with the same word as the verb

(the English verb form), or it is not aligned at all. In this case, a rule can be stated

adding the relation between the aspect marker and the verb to the Chinese annotation

in a regular way.

The work reviewed in this section illustrates the variety of cross-linguistic issues which

can be addressed on the basis of data automatically extracted from parallel corpora.

Despite the limitations discussed in Section 3.1.2, parallel corpora, in combination with

automatic word alignment, provide a new rich resource for studying various aspects of

cross-linguistic variation.

Note that, in addition to word alignment, which is common to all studies, extracting

linguistic information from parallel corpora requires other kinds of automatic processing.

If we want to extract all the instances of a certain verb in a corpus, we need to make sure

that, when we look for the verb go, for example, we obtain the instances of goes, went,

gone, and going as well. This means that the corpus needs to be lemmatised. The corpus

also needs to be morphologically tagged, so that we know that our extracted instances

are all verbs, and not some other categories. For example, we need to make sure that

the extracted instances do not include cases such as have a go, where go is a noun. If

we want to count how many times a verb is used as transitive and how many times

as intransitive, the corpus needs to be syntactically parsed. The details of linguistic

100

3.3. Statistical analysis

processing used in our experiments are explained in the methodological sections of each

case study separately, because the approaches to the studied constructions required

different linguistic analyses.


Once the aligned instances of verbs that interest us are extracted from parallel cor-

pora, we analyse them using various statistical methods. Statistical analysis allows us

to identify tendencies in the use of verbs which are relevant for studying their lexical

representation. In this section, we lay out the methods used in our studies together with

the technical background necessary for following the discussion in the dissertation. The

survey of the notions in the technical background relies mostly on two sources, Baayen

(2008) and Upton and Cook (1996).

3.3.1. Summary tables

In all the three case studies in this dissertation, observations are stored as two kinds

of variables. We distinguish between instance observations which refer to the charac-

teristics of use of verbs at the token level, and type observations, which refer to the

properties of verbs as separate entries in the lexicon. As an illustration of the two kinds

of data extracted from corpora, simple artificial examples are shown in Tables 3.1 and

3.2. Instance variables contain the information about each occurrence of a verb in the

corpus. Table 3.1, for example, contains two variables: the morphological form of the

verb in the given instances and its syntactic realisation (whether it is used as transi-

tive or not). Type variables contain the information that is relevant for lexical items

at the type level. Frequency in the corpus shown in Table 3.2 is typically the kind of

information that applies to types.

Simple tables that list the values of the variables usually do not help much in spotting

interesting patterns; individual observations are of little interest for a statistical analysis.

What is more interesting is the relationship between the values in two or more variables.

101


Instance ID Morph Transitive1 past no2 present yes3 present yes4 past no

Table 3.1.: Examples of instancevariables

Verb Frequencystop 236drive 75hide 13sleep 9

Table 3.2.: Examples of typevariables

For instance, a question that immediately comes to mind looking at Table 3.1 is whether

the verb tense somehow influences the transitivity of a verb use or the other way around.

The observations listed in the table suggest that there is a pattern: the instances that

are in the present tense are also transitive and those that are in the past simple tense

are intransitive.

A simple way to look up the relations between the values of two or more variables is

to construct a contingency table which shows the number of joint occurrences of all the

pairs of values. Table 3.3 is a contingency table which summaries the observations listed

in Table 3.1. The benefits of contingency tables might not look that obvious on such

a small data set, but as soon as the number of observations becomes greater than ten,

such summaries are necessary. The more variables and possible values the harder to see

the relationships in simple tables.

simple past presenttransitive 0 2

not transitive 2 0

Table 3.3.: A simple contingency table summarising the instance variables

Of course, the pattern that seems to be present in Table 3.1 might be due to chance

and not to a true relationship between the two variables. This is a possibility that can

never be completely excluded. Assessing the probability that patterns in observations

are due to chance is thus one of the core issues in statistics. If the probability is very

low (usually the threshold is set to p < 0.05), the pattern is significant.

102


What assessing this probability easier in a general sense is a greater number of observa-

tions. Misleading patterns occur much easier in small than in big samples. On the other

hand, true relationships can also go unnoticed in small samples. This is why we insist on

collecting and analysing large data sets. Patterns that are obvious in large samples are

very likely to be statistically significant. But one should bear in mind that, no matter

how large our collections of observations are, they still represent just small samples of

the phenomena that are generally possible in language. Their analysis makes sense only

in the context of statistical inference.

3.3.2. Statistical inference and modelling

The main purpose of a statistic analysis, as underlined by Upton and Cook (1996), is not

describing observed phenomena, but making predictions about unobserved phenomena

on the basis of a set of observations. The pattern that we observe in our toy example

in Table 3.1 is not very interesting by itself. It would become much more interesting

if we could conclude on the basis of it what the morphological form and the syntactic

realisation of every new instance of the verb will be.

Good predictions rely on good understanding of the relationships between the values

of variables. If the relationships are understood well enough, we can identify a general

rule that generates and at the same time explains the observations in the sample. As

an illustration of these notions, we adapt a simple example composed by Abney (2011).

Consider the variables recorded in (3.6).

(3.6)

t d

1 0.5

1 1

2 2

3 ?

4 7

The column t specifies the time at which an observation is made. The column d specifies

the values recorded: the distance travelled by a ball rolling down an inclined plane.

103


There are two measurements for the time t = 1 (0.5 and 1). There is no observation at

t = 3.

d = 2t−1 d = 12t2 (3.7)

In this case, the rules that generate the observed sequences can be stated as formulas.

Two possible generalisations are given in (3.7). They both capture the sequence of

observations only partially. Even if we were allowed to choose the values which are

easier to explain (which we are not), and to ignore the value 0.5 at t = 1, the formula

on the left hand side does not predict the value 7 at t = 4, but 8. The value 0.5 at t = 1

would suit better the formula on the right hand side, but neither this formula explains

the value at t = 4.

If we knew all the distances at all time points with certainty and if these values fol-

lowed a perfectly regular pattern, this pattern could be described in terms of a single

generalisation which would have no exceptions and on the basis of which any distance

at any time could be predicted, including the missing value at t = 3. Such reasoning is

common to all inductive scientific methods. Certain facts are, however, rare in science

and observations are hardly ever explainable with a single powerful generalisation. The

situation usually resembles much more the example in (3.6): we do not know the facts

for sure and we cannot explain them entirely. This perhaps applies especially to linguis-

tic phenomena, which are essentially subject to interpretation. Statistical inference is

a way to make predictions taking into consideration the uncertainty and the limits of

explanation.

Statistical predictions are formulated as the probability that a certain variable will take

a certain value (or that it will be situated within a certain range of values) under certain

conditions. The probability is usually assessed as the relative frequency of the variable

values in a sample of the studied phenomena. For example, the sample of observations

in Table 3.1 contains four observations for the morphological form variable and four for

the syntactic realisation variable. Out of four morphological forms, two are the simple

present tense and two are the past tense. The probability that the next verb is in

the simple present tense is thus equal to the probability that it is in the past tense,

104


Figure 3.2.: Probability distributions of the morphological forms and syntactic realisa-tions of the example instances.

p = 24

= 0.5. The same calculation can be done for the other variable resulting in the

same probabilities.

Assigning a probability to all possible values of a variable results in a probability distri-

bution, which can be graphically represented with a histogram. Histograms representing

probability distributions of the variables in Table 3.1 are given in Figure 3.2. Figure 3.3

shows the probability distribution of the data in Table 3.2 in two cases. The histogram

on the left hand side shows the probability distribution over verbs (how likely is an

occurrence of the verb), and the one on the right hand side shows the probability distri-

bution over the frequency values (how likely each frequency value is). For the reason of

simplicity, we assume in both cases that the lexical inventory consists of only these four

verbs.

As we can see in Figure 3.3, the shape of the distributions can be very different. The

notion of the shape of a distribution does not concern only the visual representation of

data, it is very important for inference. The patterns which are observed in the sample

can be generalised to a bigger population only if we can assume that the shape of the

105


Figure 3.3.: Probability distributions of the example verbs and their frequency.

probability distribution in the unobserved values is the same as in the values observed in

the sample. Moreover, generalisations are often only possible if we can assume a specific

shape of the probability distribution.

The shape of a distribution is determined by the values of a certain number of parameters.

The most typical examples of such parameters are the mean value and the standard

deviation (showing how much the values deviate from the mean value). There can be

other parameters depending on what kind of variation in the values of variables needs

to be captured.

The normal distribution, illustrated in Figure 3.4, is frequently referred to in science, as

many statistical tests require this particular distribution. It is characterised as symmetric

because the values around the mean value are at the same time the most probable values.

The values which are lower and higher than the mean value are equally probable, with

the probability decreasing as they are further away from the mean. Many quantitative

variables follow this pattern. A typical example is people’s height: most people are of a

medium height, while extremely tall and extremely short people are very rare. Frequency

of words in texts, for example, does not follow this pattern. There are usually only

106


Figure 3.4.: A general graphical representation of the normal distribution.

few words that are extremely frequent, but there are many words with extremely low

frequency, many more than those with medium frequency. (Our artificial example in

Table 3.2 and on the left hand side in Figure 3.3 illustrates this tendency as much as

this is possible with only four examples.) Since standard formulas for statistical tests

usually assume the normal distribution, one has to be careful when applying them to

linguistic data.

An example of a standard test which is very frequently used and which requires that

the probability distribution over values is normal is the t-test. This test is a formula

that uses the parameters of probability distributions in two samples to calculate the

probability that two samples belong to the same larger population. It is very frequently

used because it is often important to show that two samples do not belong to the same

larger population, that is that they are significantly different. In one of our case studies,

the t-test is used to show that two samples belong, in fact, to the same population.

107


As already mentioned above, real predictions based on statistical inference rarely con-

cern only one single variable. What is usually studied in statistical approaches are the

relationships between the values of two or more variables. By observing the values in

the sample, we try to determine whether the values of one variable (called dependent,)

depend on other, independent, variables. If we can determine that the values in the

dependent variable systematically increase as the values in the independent variables

increase, then we say that there is a positive correlation between the variables. If the

changes in the values in the dependent and independent variables are consistent but

in the opposite direction (increasing in one and decreasing in the other), we say that

there is a negative correlation. For example, people’s height and weight are positively

correlated: taller people generally have more weight, despite the fact that this is not

always the case. There are a number of statistical tests which measure the strength and

the significance of correlation between two variables.

The notion of correlation is fundamental to constructing statistical models. If there is

a correlation between an independent and a dependent variable and if the values of

both variables are normally distributed, then the values in the dependent variable can

be predicted from the values in the independent variables. In this case, we say that

the variation in the dependent variable is explained by the variation in the independent

variable. The purpose of statistical models is to predict values of one variable on the basis

of information contained in other variables. They model a piece of reality in terms of a

set of independent variables, potential predictors, one dependent variable, and precisely

described relationships between them. The prediction is usually based on a regression

analysis which shows to what degree the variation in the dependent variable is explained

by each factor represented with an independent variable.

3.3.3. Bayesian modelling

An alternative approach to predicting values in one variable on the basis of values in

other variables is Bayesian modelling. In this framework, the probability of some variable

taking a certain value is assessed in terms of a prior and a posterior probability. The

prior probability represents our general knowledge about some domain before learning a

new piece of information about it. The posterior probability is the result of combining

108


Variable Value Notation Probabilitya burglary in the given neighbour-hood

happens p(b) 0.014

the alarm if there is a burglary activated p(a|b) 0.75the alarm if there is no burglary activated p(a|¬b) 0.1

a burglary in the neighbourhoodif the alarm is activated

happens p(b|a) ?

Table 3.4.: An example of data summary in Bayesian modelling

the prior probability with some newly acquired knowledge. Probability updating is

formulated as conditional probability, which can be calculated from joint probability (the

probability that the variable A takes the value a and that the variable B at the same

time takes the value b) using the general conditional probability rule given in (3.8).

P (A|B) =P (A,B)

P (B)(3.8)

Bayesian modelling is based on the assumption that our knowledge about the world

is formed in a sequence of updating steps and that it can be expressed in terms of

conditional probabilities, as illustrated in Table 3.4. The example, based on Silver

(2012), concerns assessing the probability that a burglary actually took place if an

alarm is activated. In assessing this probability, we rely on several facts (listed in Table

3.4). From previous experience, we know that the probability of a burglary in the given

neighbourhood is 0.014. This is the prior probability of a burglary in our example. We

also have an assessment on how efficiently the alarm detects the burglary: it gives a

positive signal in 75% of cases of an actual burglary, and in 10% cases where there is

no burglary. We combine this knowledge by applying the equation in (3.9), known as

Bayes’ rule, which is derived from the conditional probability rule (3.8) applying the

commutative law.

P (A|B) =P (B|A) · P (A)

P (B)(3.9)

109


When we replace general symbols in (3.9) with the notation from our data summary,

we obtain the equation in (3.10). Replacing the terms with actual probability given

in Table 3.4, as in (3.11) we obtain the answer to the initial question: the probability

that a burglary took place when the alarm is activated is around 0.1, which is still low

considering that the signal from the alarm is positive.

p(b|a) =p(a|b) · p(b)

p(a)(3.10)

p(b|a) =0.75 · 0.014

0.1091= 0.096 (3.11)

Note that the term p(a) was not listed in the table. It is calculated from the conditional

probabilities which are available. As shown in (3.12), the probability that the alarm is

activated is first expressed as the sum of two joint probabilities: the probability that the

alarm is activated and there is a burglary and the probability that the alarm is activated

and there is no burglary (the probability of the complement set of values). Since the

two joint probabilities are not listed in our data, we calculate them from the conditional

probabilities which are known applying the rule in (3.8). The term p(¬b), which is

required for this calculation is given by p(b). Since these two cases are complementary,

their probability sums to 1, which yields p(¬b) = 1− p(b) = 0.986.

p(a) = p(a, b) + p(a,¬b)= p(a|b) · p(b) + p(a,¬b) · p(¬b)= 0.75 · 0.014 + 0.1 · 0.986

= 0.0105 + 0.0986

= 0.1091

(3.12)

These relatively simple calculations provide a formal framework for updating the prior

probability after having encountered new evidence related to the question that is inves-

tigated. In our example, the prior probability of a burglary in the given neighbourhood

is updated having learnt that the alarm had been activated. This updating is performed

110


taking into consideration the uncertainty that is inherent to the knowledge about the

phenomenon at each step.

An advantage of Bayesian modelling compared to the “standard” statistical inference

laid out in Section 3.3.2 is that it offers a more straightforward mechanism for combining

evidence. In the standard approach, the influence of all potential predictors on the pre-

dicted variable is assessed directly. The explanations from predictors can be combined

in a linear or weighted fashion, but not hierarchically. Contrary to this, Bayesian calcu-

lations can be applied recursively: once a posterior probability is calculated, it can be

used as a prior for some other posterior probability. For example, the prior probability

of a burglary in a particular neighbourhood, which is used in the calculations above,

could have been calculated as a posterior probability relating the chance of a burglary in

general to the relationship between some characteristics of a particular neighbourhood

and its proneness to burglaries.

Another advantage of Bayesian modelling is that it assumes no particular parameters of

probability distributions over the values of variables. The accent in Bayesian modelling

is on combining the probabilities, while their origin is less important. The probability

assessments can be expressions of intuitive (expert) knowledge, of previous experience,

or of a relative frequency in a sample. The calculations yielding new assessments apply

to any kind of probability distributions over the values as long as the probability of all

the values sums to 1 (like the probability that a burglary happens and the probability

that it does not happen in our example).

Both underlined advantages are especially important in the context of modelling linguis-

tic phenomena. The recursive nature of Bayesian models makes them a well-adapted

framework for a statistical approach to linguistic structures, which are, according to the

majority of theoretical accounts, recursive. The fact that the inference in this approach

does not depend on a particular probability distribution (notably, on the normal distribu-

tion) is important because linguistic data are often associated with unusual distributions

for which it is hard to define a small set of appropriate parameters.

These advantages, however, come at a cost. Stepping out of the standard statistical

inference framework makes evaluating the predictions in Bayesian modelling harder.

Good predictions in the traditional statistical modelling are guaranteed by the notion

111


of statistical significance. If a statistically significant effect of a predictor on a predicted

variable is identified, the predictions based on this relationship can be expected to be

correct in the majority of cases. The notion of statistical significance is not incorporated

in the predictions in Bayesian modelling. The quality of predictions has to be evaluated

in another way, usually by measuring the rate of successful predictions.

In this dissertation, both approaches are used. We apply standard tests in the situation

where we can assume the normal probability distribution over the values of a variable and

where the hierarchical relationships between the components of a model are not complex.

Otherwise, we formalise our generalisations in terms of Bayesian models and we test the

predictions comparing the predicted and the actual values on a sample of test examples.

The generalisations which are addressed by the models concern the relationship between

semantic properties of verbs and the observable formal properties of their realisations in

texts. We explain the variation in the verb instances by the variation in their semantic

properties.

3.4. Machine learning techniques

Statistical models which are proposed in this dissertation are developed by combing

theoretical insight with some standard machine learning techniques. Theoretical analysis

results in a small number of variables which define the studied domain. It also provides

the hypotheses about the dependency relationship between the variables, but the exact

numerical values of the relationships between the values of the variables are acquired

automatically from the data set.

Automatic acquisition of generalisations from data is studied in the domain of machine

learning. Using the general terminology of learning, the data which machine learning

algorithms take as input are regarded as experience. A computer program is said to learn

from experience if its performance at some task improves with the experience, that is

by observing the data. In our experiments, we assume that the machine learning task

is defined as classification. The notions used in the section are mostly based on three

sources, Mitchell (1997), Russell and Norvig (2010), and Witten and Frank (2005).

112


There are two main approaches that can be taken in inferring the relationships between

the values of variables: supervised and unsupervised learning. In this section, we first

illustrate the two approaches by describing standard algorithms which are most widely

used. We then show how the two approaches are implemented with Bayesian models in

this dissertation.

3.4.1. Supervised learning

In the supervised learning setting, the training data include the information about the

values in the predicted variable, which we call the target variable. To illustrate these

notions, we adapt an example constructed by Russell and Norvig (2010) (see the data

summary in Table 3.5). Suppose we are at work and we receive a message that our

neighbour Mary called some time ago. We want to assess the probability that this call

means that there was a burglary at our place. We have an old-fashion alarm that rings

when some shock is detected, but we cannot hear the ringing when we are away from

home. So we ask our neighbours Mary and John to call us if they hear the alarm. The

day when we receive a call from Mary, we did not hear from John. We should also bear

in mind that Mary could have called for some other reason. Also the alarm could have

been activated by some other shock, not burglary (like an earthquake, for example). The

question that we ask is: What is the probability that there was a burglary, given that

Mary called, John did not call and there was no earthquake, p(b = yes|m = yes, j =

no, e = no) (the bottom row in Table 3.5)?

In assessing the probability, we look up the records of the last ten cases when one of our

neighbours called us at work (the other rows in Table 3.5). What we want to find is the

same situations in the past and to see whether a burglary actually happened in these

cases. We find that only one of the previous situations (row seven) was the same and

that a burglary actually happened then. However, we are still not convinced because in

the majority of our records there was no burglary.

To use all the data available, we look how the burglary was related to each of the values

separately and then we recompose the probability for the case in queastion. To illustrate

113


Mary calls John calls Earthquake Burglary1 yes yes no yes2 yes yes no yes3 yes no yes no4 no yes no no5 no yes no no6 no yes no no7 yes no no yes8 no yes no no9 no yes no no

10 yes yes yes noQ yes no no ?

Table 3.5.: An example of a data record suitable for supervised machine learning

how this can be done, we describe two algorithms which are usually regarded as rather

simple, but often well performing.

The Naıve Bayes algorithm decomposes the records assuming that the values of the three

predictors are mutually independent. The term naıve in the name of the algorithm refers

to the fact that the variables are usually not independent in reality, but that the potential

dependencies are ignored. With the assumed independence, the probability which we

look for can be expressed as the product of individual conditional probabilities, as shown

in (3.13). Since our task is classification, we look for the probability of a particular class

cj based on the values of predictor variables, which are attributes a1,...,n of each instance

of a class.

P (a1, a2, a3, ...an|cj) ≈∏i

P (ai|cj) (3.13)

We calculate the most probable class for a given set of values of attributes by applying

Bayes’ rule in (3.9) repeated here as (3.14), which gives the general formula in (3.15),

where the 1z

is a constant when the values of the attributes are known, as in our example.

When we apply the general classification formula to our data, we obtain (3.16).

114


P (A|B) =P (B|A) · P (A)

P (B)(3.14)

P (cj|ai, ...an) ≈ 1

zP (cj)

n∏i

P (ai|cj) (3.15)

p(b = yes|m = yes, j = no, e = no) ≈ 1z· p(b = yes) · p(m = yes|b = yes)

·p(j = no|b = yes) · p(e = no|b = yes) (3.16)

With the separated conditional probabilities, we can use more records to estimate each

factor of the product. For example, applying the conditional probability rule in (3.8),

we can calculate:

p(m = yes|b = yes) =p(m = yes, b = yes)

p(b = yes)=

3

3

Burglary actually happened in three of our ten records and in all three of them, Mary

called. The score would be different for John, since he failed to call in one of the three

cases.

In deciding whether to classify the current situation as burglary or not, we calculate

the product of all the conditional probabilities for both potential values of the target

variable, we multiply this product with the probability of the value of the target variable,

and then we select the higher probability. The constant can be omitted because it is the

same for both classes. In this particular example, the calculation would give:

p(b = yes|m = yes, j = no, e = no) ≈ 3

10· 3

3· 1

3· 3

3=

1

10(3.17)

p(b = no|m = yes, j = no, e = no) ≈ 7

10· 2

7· 1

7· 5

7=

1

49(3.18)

115


Mary calls?

Yes.

Earthquake?

Yes.

No burglary.

No.

Burglary!

No.

No burglary.

Figure 3.5.: An example of a decision tree

Since 110> 1

49, the final decision should be to classify Mary’s call as signalling a bur-

glary.

Applying the decision tree algorithm to the same data set, we proceed by querying each

variable in the order of informativeness, as shown in Figure 3.5. We first determine

that Mary called. If she had not called there would have been no reason to worry. But

Mary did call in this case, so then we check if there was an earthquake immediately

preceding Mary’s call. If there was an earthquake, there is no need to worry; it was the

earthquake that probably activated the alarm, which made Mary call. But if there was

no earthquake, which is the case in the current situation, then we should better hurry

home, because there was a burglary.

The decision tree which brought us to this conclusion is constructed on the basis of

the same records which were used for assessing the probabilities for the naıve Bayes

algorithm (Table 3.5). In deciding which variable should be in the root of the tree, we

look up the joint distributions of values combining each predictor with the target variable

separately. This procedure results in groupings shown in the upper part of Table 3.6

(Step 1). We compare the resulting divisions to identify the most discriminative variable.

The variable which gives the “purest” groups is the most discriminative. In our example,

two variables give entirely pure goupings. We can see in Table 3.6 that every time the

116


value of Marry calls is “no”, the value of the target variable is also “no”. But also every

time the value of Earthquake is “yes”, the value of the target variable is “no”. Since

they give pure classes, these two variables are candidates for the most discriminative

variable. Mary calls wins because it gives the bigger pure group and also because the

size of the two resulting groups, which depends on the distribution of the values in the

variable, is more balanced (there are five occurrences of each value).

If all the values of the target variable were the same when the value of Mary calls is

“yes”, we could stop at this point, ignore the other two variables and predict the target

variable only from the values of Mary calls. This is, however, not the case, so we need

to continue constructing the tree by looking up the combinations of values of Mary calls

with the other two variables. The aim here is to see if some of these combinations will

result in pure groupings of the values in the target variable. The resulting groupings of

the second step are shown in the bottom part of Table 3.6 (Step 2).

We can see that the combination of values of Mary calls and Earthquake divides the set

of values of the target variable into entirely pure classes (for both values of Mary calls).

Since all the groupings of the values in the target variable are pure at this point, the tree

is completed. We can ignore the variable John calls because it provides no information

about whether there was a burglary or not.

We have used the notion of the purity of a class in an intuitive way so far. We have

considered the classes purer if they contain more of the same kind of items. For example,

the group of values of the target variable associated with the value “yes” of the variable

John calls in the upper part of Table 3.6 is “purer” than the group of values associated

with the value “no” of the same variable because the proportion of the same items is

bigger in the first group (five out of seven) than in the second (two out of three). The

same principle applies when working with large data set, but the purity of classes has

to be measured at each step as it is hard to assess larger classes intuitively.

The measure that is most commonly used to assess the purity of classes is entropy,

which is calculated from the probability distribution of a variable, using the formula in

(3.19), where S denotes the variable for which we measure the entropy and c denotes the

number of possible values of the variable. It shows the degree to which the probabilities

117


Step 1 Mary calls John calls Earthquakeyes no yes no yes no

Burglary yes no yes no no yesyes no yes no no yesno no no yes noyes no no nono no no no

no yesno no

no

Step 2 Mary calls Mary callsyes no yes no

John calls John calls Earthquake Earthquakeyes no yes no yes no yes no

Burglary yes no no no yes noyes yes no no yes nono yes no yes no

no no

Table 3.6.: Grouping values for training a decision tree

of individual values vary. The more similar the probabilities of the values, the higher

entropy. If one value is much more likely than the other, entropy is going to be low.

Entropy(S) =c∑

i=1

− pilog2pi (3.19)

As an illustration, we calculate the entropy of the set of Burglary values observed in the

training data. There are two possible values: “yes” occurs three times, and “no” occurs

seven times.

118


Entropy(B) = − 310log2

310− 7

10log2

710

= −(0.3 · −1.74)− (0.7 · −0.51)

= 0.52 + 0.36

= 0.88

(3.20)

To choose the attribute which should be put in the root of the decision tree, we compare

the entropy of the starting set of values with the entropy of the subsets of values of the

target variable which are associated with each value of each attribute (the columns in

the upper part of Table 3.6). The variable that is considered the most discriminative at

each node in constructing a decision tree is the one which reduces the most the entropy

of the target variable. The measure which is most commonly used for this comparison

is called information gain. It is calculated using the formula in (3.21), where A is the

attribute which is considered and |Sv| is the subset of S for which the value of A is v.

Gain(S,A) = Entropy(S)−∑

v∈V alues(A)

|Sv||S|

Entropy(Sv) (3.21)

As an illustration, we calculate the information gain of the attribute Marycalls in our

example:

GainB,M = Entropy(B)∑

v∈V alues(M)|Bv ||B| Entropy(Bv)

= Entropy(B)− |BM=yes||B| Entropy(BM=yes)− |BM=no|

|B| Entropy(BM=no)

= 0.88− 510· 0.97− 5

10· 0

= 0.88− 0.48− 0

= 0.40

(3.22)

The same calculations are performed for the other two attributes and the one which

provides the highest information gain is taken as the first split attribute. The calculations

are performed recursively until the entropy of all resulting subsets is 0. Note that

the entropy of BM=no in our example is 0 because all the values in this subset are

the same. In this case, the recursive calculations are performed only for the subset

119


Mary calls Earthquake Burglary1 yes no ?2 yes no ?3 yes yes ?4 no no ?5 no no ?6 no no ?7 yes no ?8 no no ?9 no no ?

10 yes yes ?Q yes no ?

Table 3.7.: An example of a data record suitable for supervised machine learning

BM=yes. In practice, the programs that implement the decision tree algorithm work

with some additional constraints, but we do not discuss these issues further because

such a discussion would exceed the scope of this survey.

3.4.2. Unsupervised learning

In the unsupervised learning setting, the values of the target variable are not known

in the training data. The task of deciding what class to assign to a particular case

resembles the data summary in Table 3.7. The data in Table 3.7 represent basically

the same records as in Table 3.5, with the variable John calls omitted for the reason of

simplicity. The question marks in the last column represent the fact that the values of

the target variable are not recorded. However, we can assume that such a variable exists

and that its values can be explained by the values of the other, known variables. Such

variables are called hidden variables.

In principle, hidden variables are not necessarily the target variables. Any variable in a

model can be regarded as hidden. In some models, the values of the target variable itself

are known in training, but they are assumed to be influenced by some other variable with

an unknown probability distribution. In this case, the learning setting is supervised,

120


but estimating the probability distribution of the hidden variable requires a special

approach.

In this subsection, we describe the expectation-maximisation algorithm, which is often

used in assessing the probability distribution of a hidden variable, regardless of whether

it is a target variable. It is a general algorithm which has been applied to many different

learning tasks. The algorithm is applied to an independently constructed model and to a

set of data in an iterative fashion. As its name suggests, it consists of two main parts. In

the expectation part, expected values in the data set are generated based on hypothesised

parameters of distributions. In the maximisation part, the hypothesised parameters of

distributions are combined with the observations in the data set and updated so that they

are more consistent with the observed data. As a result, the parameters of distributions

of both observed and unobserved variables are consistent with the observed data. The

algorithm starts with arbitrary hypothesised parameters which are combined with the

observations and updated in a number of iterations. It ends when the parameters reach

the values which are consistent with the data and they are no longer updated.

The mathematical background of the algorithm is much more complex than in the case

of the two supervised algorithms which we have introduced so far. Its precise general

mathematical formulation would exceed the scope of this survey. We thus limit the

discussion in this subsection to the particular application of the algorithm which is used

in this dissertation. To illustrate the functioning of the algorithm, we use the same model

and the same data set which were used for the naıve Bayes algorithm in Section 3.4.1,

modified as shown in Table 3.7. With the variable John calls omitted (for simplicity),

the model is formulated as in (3.23).

p(m, e, b) = p(b) · p(m|b) · p(e|b) (3.23)

The model formulation which we use in this example is more general than in the previous

calculations. Instead of specifying concrete values, we refer to any value that a variable

can take. Thus the small letter m stands for both Mary calls = “yes” and Mary calls

= “no”, e stands for both values of Earthquake, and b for both values of Burglary. Note

also that the value on the left side of the equation is not a conditional, but a joint

121


Iteration 1 Iteration 2m e b p(b) p(m|b) p(e|b) p(M) Cc p(b) p(m|b) p(e|b) p(M) Cc

y y y 0.4 0.4 0.4 0.064 0.8 0.4 0.5 0.2 0.04 0.8y y n 0.6 0.4 0.4 0.096 1.2 0.6 0.5 0.2 0.06 1.2y n y 0.4 0.4 0.6 0.096 1.2 0.4 0.5 0.8 0.16 1.2y n n 0.6 0.4 0.6 0.144 1.8 0.6 0.5 0.8 0.24 1.8n y y 0.4 0.6 0.4 0.096 - 0.4 0.5 0.2 0.04 -n y n 0.6 0.6 0.4 0.144 - 0.6 0.5 0.2 0.06 -n n y 0.4 0.6 0.6 0.144 2 0.4 0.5 0.8 0.16 2n n n 0.6 0.6 0.6 0.216 3 0.6 0.5 0.8 0.24 3

Table 3.8.: An example of probability estimation using the expectation-maximisationalgorithm

probability, which is also a more general case (knowing the joint probability of a set of

values allows one to calculate several related conditional probabilities).

The more general formulation is needed for this example because the expectation-

maximisation algorithm explores all possible combinations of values. In our example,

there are three variables, each with two possible values. The number of possible combi-

nations of values is 23 = 8. They are all listed in the first three columns in Table 3.8,

which shows two iterations of the algorithm assuming the model in (3.23) and the data

set in Table 3.7.

The first step of the algorithm is the initialisation of the model. In this step, the

probability distribution of all variables is determined in an arbitrary way, regardless

of the frequency of certain values in the training sample. For example, we assign the

probability 0.4 to all values “yes” of all variables (regardless of whether its probability

is conditional or prior) and 0.6 to all values “no” of all variables. This initialisation

reflects our general belief about which events are more likely and which ones are less

likely. Note, however, that the initialisation step is arbitrary and relating it to some

existing belief does not guarantee a better result.

The initial (arbitrary) probability of each factor of the model is shown under Iteration

1 in Table 3.8. The probability of the whole model (column p(M) in the table) is

122


calculated by multiplying the factors, as shown for the first two cases in (3.24) and

(3.25) respectively. The probability of the model in each case is then combined with the

counts observed in Table 3.7 to distribute the counts to different cases. The distributed

counts are called complete counts (Ccounts in the formulas, Cs columns in Table 3.8)

as opposed to the incomplete counts which are available in the data set. For example,

what we can see in Table 3.7 is that there were two cases where both Mary called and

there was an earthquake (F (y, y, ∗) in the formula, the asterisk stands for any value of

the third variable), but we do not know whether there was a burglary in these cases. We

apply the formula shown in (3.24) and (3.25) to assign the count of 0.8 to the first case

(where there is a burglary) and the count 1.2 to the second case (where there was no

burglary). Note that the counts are fractional, which would not be possible in reality,

but this is acceptable because they are only an intermediate step in calculating the

probability of each factor of the model. Applying the formula gives complete counts for

all the cases, as shown in the column Cc under Iteration 1 in Table 3.8.

b=y, m=y, e=y:

p(M) = p(b = y) · p(m = y|b = y) · p(e = y|b = y)

= 0.4 · 0.4 · 0.4 = 0.064

Ccounts =F (y,y,∗)·p(M(y,y,y))

p(M(y,y,y))+p(M(y,y,n))= 2·0.064

0.064+0.096= 0.128

0.16= 0.8

p(b = y) = F (b=y)Total

= 0.8+1.2+0+210

= 410

= 0.4

p(m = y|b = y) = F (m=y,b=y)F (b=y)

= 0.8+1.24

= 0.5

p(e = y|b = y) = F (e=y,b=y)F (b=y)

= 0.8+04

= 0.2

New P(M) = 0.4 · 0.5 · 0.2 = 0.04

(3.24)

123


b=n, m=y, e=y:

p(M) = p(b = n) · p(m = y|b = n) · p(e = y|b = n)

= 0.6 · 0.4 · 0.4 = 0.096

Ccounts =F (y,y,∗)·p(M(y,y,n))

p(M(y,y,y))+p(M(y,y,n))= 2·0.096

0.064+0.096= 0.192

0.16= 1.2

p(b = n) = F (b=n)Total

= 1.2+1.8+0+310

= 610

= 0.6

p(m = y|b = n) = F (m=y,b=n)F (b=n)

= 1.2+1.86

= 0.5

p(e = y|b = n) = F (e=y,b=n)F (b=n)

= 1.2+06

= 0.2

New P(M) = 0.6 · 0.5 · 0.2 = 0.06

(3.25)

To update the probability of each factor of the model, we sum up the counts for each

relevant case and calculate the conditional probability applying the conditional proba-

bility rule (3.8), as shown for the first two cases in (3.24) and (3.25). All the counts

which are added up can be looked up in the corresponding cells under Iteration 1 in

Table 3.8, and all the resulting updated probabilities of each factor of the model in each

case are listed under Iteration 2.

In the next step, we calculate the probability of the model by multiplying the updated

probabilities of the factors. We then calculate new complete counts (the Cc column

under Iteration 2 in Table 3.8) using the updated model probability and then use the

new counts to update again the probability of the factors of the model. We then repeat

applying and updating the model till the probabilities of the factors of the model converge

to the true probabilities. The convergence is not guaranteed in all the cases, but if the

patterns in the data are clear enough, it is very likely.

Looking at the values in Table 3.8, we can see that the initial arbitrary probabilities

of the models have changed when combined with the information about the incomplete

counts. In both cases which are of interest in our example (the cases (y, n, y) and (y,

124


n, n), the probability of the model has increased. The probability of no burglary is

still higher than the probability of burglary, which does not correspond to the results

of supervised learning in Section 3.4.1. However, the ranking of the two models would

change if there were more instance to learn from.

Unsupervised learning is harder than supervised learning because crucial information

about the values of the target variable is not available for training. However, it is

increasingly used in natural language processing because linguistic data sets with known

target variables, such as manually annotated corpora presented in Chapter 2, Section

2.2.2, are hard to construct. Another reason why unsupervised learning is seen as an

attractive framework for approaching linguistic phenomena is the fact that it allows using

corpus data for discovering new structures, not pre-defined in linguistic annotation, such

as in the experiments on grammar induction by Klein (2005).

3.4.3. Learning with Bayesian Networks

Since the main purpose of the models proposed in our case studies is representing gen-

eralisations about the structure of language, the accent is not as much on assessing the

probabilities as it is on the structure of the relationships between the variables. We

use rather basic learning methods for training the models on a set of data extracted

from corpora, putting more complexity on the structure of the models. To represent

the hierarchical relationships between the variables, we formulate our models in terms

of Bayesian networks.

A Bayesian network is a directed acyclic graph where the nodes represent the variables

of a model and the edges represent the dependency relationships between the variables.

A Bayesian network would be very useful, for example, if we wanted to add the variable

Alarm to the model discussed in Section 3.4.1. Although we know that Mary’s and

John’s calls depend, to a certain degree, on whether they have heard the alarm, this

dependence is only implicitly present in the data set. A graph such as the one in Figure

3.6 can be used to represent the role of the alarm explicitly.

The edges in the graph show that the alarm can be caused by an earthquake or by a

burglary, and also that it causes John and Mary to call. Each edge is associated with a

125


Earthquake Burglary

Alarm

Mary calls John calls

Figure 3.6.: An example of a Bayesian network

conditional probability distribution showing how the two variables which are connected

with it are related. For example, we can specify the probability of the alarm being

activated by an earthquake as p(a|e) = 0.7, which also specifies p(¬a|e) = 0.3. We

can specify the probability of a burglary activating the alarm as higher, for example

as p(a|b) = 0.9 and p(¬a|b) = 0.1. Such conditional probabilities are specified for each

node and each edge. They can be estimated on the basis of intuition or on the basis of

training on a set of examples using machine learning algorithms. Some of the variables

can be regarded as hidden and their distribution estimated using approaches such as the

one described in Section 3.4.2.

The probability of the whole model represented in Figure 3.6 is given in (3.26):

p(e, b, a,m, j) = p(e) · p(b) · p(a|e) · p(a|b) · p(m|a) · (j|a) (3.26)

The decomposition of the model into the factors is based on the notion of conditional

independence, which allows us to reduce the complexity of the potential dependencies

between the variables avoiding at the same time oversimplifications, such as for example

the independence assumption used in the naıve Bayes algorithm (see Section 3.4.1). Note,

for example, that the variables Mary calls and John calls are not directly connected in

the graph. This represents the fact that these two variables are conditionally independent

given Alarm. If we know whether the alarm rang or not, then Mary’s and John’s call

do not depend on each other, but they both depend on the state of the alarm. Also,

126


Predicted 1 Predicted 0True 1 A BTrue 0 C D

Table 3.9.: Precision and recall matrix

note that each node in the graph depends only on the parent node (or nodes). This

means that, if we know the value of the alarm, then the calls from Mary and John are

not relevant for assessing the probability that there was a burglary. The probability of

any particular value of any variable in the network can be inferred applying Bayes’ rule

(3.9).

3.4.4. Evaluation of predictions

The success of a model in making predictions is evaluated on a test data set which

contains new instances. The values of the target variable in each instance is predicted

based on the values of predictor variables (like in the bottom rows of Tables 3.5 and

3.7). The predictions are then compared with the correct answers, usually called gold

standard, and a measure of success is calculated. The predictions of the model are

counted as correct if the predicted values are identical to the values in the gold standard.

Since a number of values can be identical to the gold standard due to chance, the success

of a model is usually defined as an improvement relative to the baseline — the result

that would be achieved by chance, or by a very simple technique.

The most commonly used measure is the F1 measure. It is the harmonic mean of two

measures: precision (p) and recall (r):

F1 =2 · (p · r)p+ r

(3.27)

Precision shows how many of the predictions made are correct (p = A(A+C)

in the matrix

in Table 3.9). Recall shows how many of the values that exist in the gold standard are

127


also predicted by the model (r = A(A+B)

).

The difference in recall and precision is important for the tasks where some instances

can be left without a response by the model (for example, these measures are typically

used in the tasks of information retrieval). Since in our experiments every instances

is given a prediction, the appropriate measure is accuracy. It is calculated using the

formula in (3.28).

Accuracy =Correct

All(3.28)

The correct predictions include true positives and true negatives, while the difference

between correct predictions and the total number of predictions includes false positives

and false negatives.

3.5. Summary

In this chapter, we have discussed two methodological issues concerning the use of par-

allel corpora for linguistic research. We have first addressed the question of why use

parallel corpora. We propose this approach as an extension of standard analysis of

cross-linguistic variation in the context of studying microparametric variation. To deal

with the linguistically irrelevant variation, which is seen as one of the main obstacles for

using parallel corpora for linguistic research, we propose collecting large data sets con-

taining maximally parallel verb instances. In addition to the methodological discussion,

we present additional arguments in favour of this approach coming from the experiments

in natural language processing, which demonstrate that automatically word-aligned cor-

pora provide a rich new resource for studying various questions related to cross-linguistic

variation. Having argued in favour of using parallel corpora, we have then discussed ap-

plying the methodology of natural language processing to address theoretical linguistic

issues by processing large data sets. As this methodology has not been commonly used

in linguistic research so far, we provide the technical background necessary for following

the presentations of our experiments in the three case studies. The introduction to the

notions in statistical inference and modelling in combination with machine learning is

128

3.5. Summary

carefully adapted specifically for the purpose of this dissertation, providing all the neces-

sary technical details, but in a way which is adapted to an audience with little experience

in these disciplines. The general methodology outlined in this chapter is applied in three

cases studies which are presented in the following chapters.

129

4. Force dynamics schemata and

cross-linguistic alignment of light

verb constructions

4.1. Introduction

Light verb constructions are special verb phrases which are identified as periphrastic

paraphrases of verbs. English expressions put the blame on, give someone a kick, take a

walk are instances of such paraphrases for the verbs blame, kick, and walk. These con-

structions are attested in many different languages representing a wide-spread linguistic

phenomenon, interesting both for theoretical and computational linguistics. They are

characterised by a special relation between the syntax and the semantics of their con-

stituents. The overall meaning of the phrase matches the meaning of the complement,

instead of matching the meaning of the head word (the verb), which is the case in typi-

cal verb phrases. Figure 4.1 illustrates the difference between regular verb phrases and

phrases headed by a light verb. Despite the same syntactic structures, the two phrases

are interpreted differently: have a yacht is about having, while have a laugh is about

laughing.

The special relation between the meaning and the structure makes light verb construc-

tions semantically non-compositional or opaque to a certain degree. The meaning of the

phrase cannot be calculated from the meaning of its constituents using general rules of

grammar. Moreover, the use of these phrases is partially conventionalised. They show

131

4. Force dynamics schemata and cross-linguistic alignment of light verb constructions

Regular VP Light verb construction

VP[syntactic features]

[semantic features]VP

[syntactic features]

[semantic features]

Verb Complement Light verb Complement

have a yaht have a laugh

Figure 4.1.: A schematic representation of the structure of a light verb constructioncompared with a typical verb phrase. The dashed arrows show the directionof projection.

some properties of idiomatic expressions, but, unlike collocations and idioms, they are

formed according to the same “semi-productive” pattern in different languages.

The semi-productive and semi-compositional nature of light verb constructions has im-

portant consequences for their cross-linguistic mappings. Consider the following exam-

ples of English constructions and their translations into German and Serbian.

(4.1) a. Mary [had a laugh]. (English)

b. Maria [lachte]. (German)

c. Marija se [na-smejala]. (Serbian)

(4.2) a. Mary [gave a talk]. (English)

b. Maria [hielt einen Vortrag]. (German)

c. Marija [je o-drzala predavanje]. (Serbian)

English expression had a laugh in (4.1a) is translated to German with a single verb

(lachte in (4.1b)). The Serbian counterpart of the English expression (nasmejala in

(4.1c)) is also a single verb, but with a prefix attached to it. By contrast, the English

expression in (4.2a) is translated with phrases both in German and in Serbian, but the

132

4.1. Introduction

heading verbs are not lexical counterparts. Unlike English gave, German hielt means

’held’, and Serbian odrzala means approximately ’held for a moment’.

Distinguishing between regular verb phrases and light verb constructions is crucial both

for constructing correct representations of sentences and for establishing cross-linguistic

mappings (Hwang et al. 2010). Moreover, one needs to distinguish between different

kinds of light verb constructions to account for the fact that they are not distributed

across languages in the same way. In some cases, cross-linguistically equivalent expres-

sions are constructions, as in (4.2), while in other cases, cross-linguistic equivalence holds

between constructions and individual lexical items, as in (4.1). However, these distinc-

tions are hard to make because there are no formal indicators which would mark the

differences either morphologically or syntactically.

The issue of distinguishing between different types of light verb constructions has been

addressed in both theoretical and computational linguistics. It has been argued that

these constructions should be seen as a continuum of verb usages with different degrees

of verbs’ lightness and different degrees of compositionality of the meaning of construc-

tions. There has been a number of proposals as to how to distinguish between different

kinds of constructions. Despite the fact that light verb constructions are headed by

several different verbs in all the studied languages (for example, take, make, have, give,

pay in English), the proposed accounts do not address the potential influence of lexical

properties of the heading verb on the overall interpretation of the construction. Regard-

ing light verbs as semantically empty or impoverished, the proposed accounts rely on the

characteristics which are common to all of them. Contrary to this, our study addresses

potential lexical differences between light verbs. We perform two experiments showing

that cross-linguistic mappings of English light verb constructions depend on the kind of

meaning of the heading light verbs. We describe the meaning in terms of force dynamics

schemata (see Chapter 2, Section 2.1.4).

The chapter consists of four main parts. In the first part, we present the questions

raised by light verb constructions and the proposed accounts which constitute the the-

oretical background of our study. We start by introducing the problem of semantic role

assignment in light verb constructions (4.2.1), which is followed by the discussion of

the proposed distinctions between different constructions (4.2.2). In the second part,

133


we present two experiments. In the first experiment (4.3.1), we examine the differences

in cross-linguistic alignments between two kinds of light verb constructions in a sample

of instances extracted from a parallel corpus based on manual word alignment. In the

second experiment (4.3.2), we evaluate automatic word alignment of the same sample of

instances which is manually analysed in the first experiment. The aim of this analysis

is to determine whether the quality of automatic alignment of light verb constructions

depends on the semantic properties of the heading light verbs. In the third part (Sec-

tion 4.4, we interpret the results of our experiments in light of the theoretical discussion

presented in the first part. We compare the findings of our study with the related work

in Section 4.5.

4.2. Theoretical background

Theoretical accounts of light verb constructions are mostly concerned with the question

of whether light verbs assign semantic roles to some constituents in a sentence or not.

While some authors argue that light verbs are functional words with no lexical content

and no predicate-argument structure, others argue that some semantic roles are assigned

by light verbs. In the following subsection, we discuss theoretical challenges posed by

light verb constructions and the proposed accounts. We then turn to the issue of semi-

compositionality and semi-productivity of the constructions.

4.2.1. Light verb constructions as complex predicates

The question of whether light verbs assign semantic roles or not is theoretically inter-

esting because it relates directly to the general theory of the relationship between the

lexical properties of verbs and the rules of phrase structure (see Chapter 2, Section

2.1.1). Note that the nouns which head the complements of light verbs, for example

look in (4.3a), are derived from verbs. Contrary to other, regular nouns, these nouns

retain the relational meaning of the original verb. For example, the noun look in (4.3a)

relates the nouns daughter and Mary in a similar way as the verb look in (4.3b). If a

light verb which heads a light verb construction (for example, took in (4.3a)) assigns

134


some semantic roles too, then there are more arguments of verbs with semantic roles

than constituents in the clause that can realise them syntactically.1 This problem is

characteristic for a range of phenomena usually called complex predicates.

(4.3) a. Mary took a look [at her daughter].

b. Mary looked [at her daughter].

In some languages, such as Urdu, (Butt and Geuder 2001), light verbs can take both

verbs and deverbal nouns as complements. In others, such as English, they only take

deverbal nouns, but these nouns can be more or less similar to the corresponding verbs.

Their form can be identical to the verb form, as it is the case with look in (4.3), or it can

be derived from a verb with a suffix (e.g. inspectV vs. inspectionN). In some cases, the

same semantic arguments of deverbal nouns and their corresponding verbs are realised

as the same syntactic complement. For example, the same prepositional phrase at her

daughter in 4.3 occurs as a complement of both the noun and the verb look. In other

cases, the same semantic argument can be differently realised in syntax (her brother vs.

to her brother in (4.4)) or it can be left unspecified (the project site vs. no complement

in (4.5)).

(4.4) a. Mary visited [her brother].

b. Mary paid a visit [to her brother].

(4.5) a. They inspected [the project site] last week.

b. They made an inspection last week.

The meaning of a deverbal noun can be more or less similar to the meaning of the

corresponding verb. Grimshaw (1990) distinguishes between event and result nominal

structures called nominals, arguing that only event nominals actually denote an action

and can take arguments. For example, the expression in (4.6a) is grammatical, while

the expression in (4.6b) is not. According to this test, the deverbal noun examination

1Note that auxiliary and modal verbs constitute a single lexical unit with a main verb. The problemof syntactic realisation of verbal arguments does not arise with these items because they are purelyfunctional words with no idiosyncratic lexical content; they do not assign to their arguments anysemantic roles that need to be interpreted.

135


refers to an activity, while exam refers to a result of an activity. In addition to this test,

Grimshaw (1990) proposes several syntactic indicators to distinguish between deverbal

nouns which refer to an activity and which are, thus, more similar to the corresponding

verbs and the nouns which refer to a result state, which is closer to the typical nominal

meaning. One of the test is the indefinite article. As illustrated in (4.7), result nominals

(4.7a) can occur in an indefinite context, while event nominals cannot ((4.7b) is not

acceptable).

(4.6) a. the examination of the papers

b. * the exam of the papers

(4.7) a. * take an examination

b. take an exam

According to this analysis, most light verb complements would be classified as result

nominals, since the indefinite article seems to be one of the characteristic determiners

in light verb constructions (see also the examples below). This characteristic, however,

does not necessarily hold in all languages.

Based on an analysis of Japanese light verb constructions, Grimshaw and Mester (1988)

provide evidence for a distinction between transparent noun phrases which are comple-

ments of the verb suru and opaque noun phrases which are complements of the verb

soseru. The former are special noun phrases which occur only as complements of light

verbs. They are described as transparent because the predicate-argument relations are

syntactically marked (by cases). The latter are more typical noun phrases which occur in

other contexts as well. They are described as opaque because the semantic relationships

in these phrases are interpreted implicitly.

According to Grimshaw and Mester (1988), English light verb constructions would all be

formed with the opaque nominals. For example, the relationship between the predicate

visit and its argument her brother is transparent in (4.4a), where visit is a verb: her

brother is theme and this relationship is syntactically expressed as the direct object.

Contrary to this, the same semantic relationship is not transparent in (4.4b), where

visit is a noun. The attachment of the prepositional phrase to her brother is ambiguous

136


(it can be attached to the light verb paid or the noun visit), and its semantic role is

interpreted implicitly (the preposition to does not encode the role theme).

Wierzbicka (1982), on the other hand, underlines the difference in meaning between the

complements of light verbs in English. For example, the meaning of the verb have in

(4.8) is contrasted to the one in (4.9-4.11). The nouns like swim in (4.8) are claimed to

be verbs “despite the fact that they combine with an indefinite article” and should be

distinguished from deverbal nouns. All the derived forms are considered to be nouns,

together with some nouns that have the same form as verbs, but whose meaning is clearly

that of a noun, such as smile in (4.9), cough in (4.10), or quarrel in (4.11). Wierzbicka

(1982), however, does not use any observable criterion or test to distinguish between the

nouns such as swim in (4.8) and the nouns such as smile, cough, quarrel in (4.9-4.11)

relying only on individual judgements.

(4.8) He had a swim.

(4.9) She has a nice smile.

(4.10) He has a nasty cough.

(4.11) They had a quarrel.

Kearns (2002) notices that the complements of light verbs in English are not “real nouns”

in some constructions, but that they are coined for light verb constructions and do not

occur freely in other nominal environments. This characteristic makes some light verb

constructions in English similar to the suru-constructions in Japanese.

The degree to which the complement of a light verb is similar to its corresponding verb

influences the overall representation of the light verb construction. The more verbal the

complement the less straightforward the assignment of semantic roles in the construction.

The more typical the noun which heads the complement, the more compositional and

regular the construction. Light verb constructions are distributed on a scale ranging

from complex predicates to near regular constructions. The variety of constructions is

discussed in the following subsection.

137


4.2.2. The diversity of light verb constructions

Several degrees of “lightness” of light verbs are illustrated by expressions in (4.12-4.16)

taken from Butt and Geuder (2001). The sequence of expressions shows the gradual

extension of the prototypical meaning of give (4.12) to its lightest use (4.16) .

(4.12) a. give him the ball

b. give the dog a bone

c. give the costumer a receipt

(4.13) a. Tom gave the children their inheritance money before he died.

b. The king gave the settlers land.

(4.14) a. give advice

b. give someone the right to do something

c. give someone information

(4.15) a. give someone emotional supported

b. give someone one’s regards

(4.16) a. give someone a kiss / a push / a punch / a nudge / a hug

b. give the car a wash, give the soup a stir

The change in the meaning of give depends on the sort of the complement. The most

prototypical variant in (4.12) involves a change in possession of the object together with

a change of its location. Having a more abstract object, or an object that does not move,

excludes the component of moving from give in (4.13). The possession is excluded with

objects that are not actually possessed, such as advice or right in (4.14), and replaced

with a more abstract component of a result state. The action of “giving” in (4.15) is

realised without “giver’s” control over the recipient’s state. Finally, the light give in

(4.16) does not describe a transfer at all, but just “the exertion of some effect on the

138


recipient”. The difference between the two groups of expressions is made by the presence

of the component of moving in (4.16a), while in (4.16b), even this is gone.

The presence of an agent (the participant that performs or causes the action described

by the verb), the completion of the action, and its “directedness” are the components

of meaning present in all the realisations. By comparing the range of uses of give in

English and its corresponding verb de in Urdu, Butt and Geuder (2001) argue that the

same components of meaning which are shared by all the illustrated uses of English give

are also the components that the English give and the Urdu de have in common.

Brugman (2001) takes a more formal approach to identifying the relevant components

of meaning on the basis of which light verb constructions can be differentiated. Instead

of analysing the properties of the nominal complements, Brugman (2001) turns to the

light verbs themselves, focusing on the English verbs give, take, and have. In an analysis

that assumes the force-dynamic schemata (Talmy 2000) (see Section 2.1.4 in Chapter

2 for more details), Brugman (2001) argues that light verbs retain the pattern of force

dynamics (or a part of it) of their prototypical (semantically specified) counterparts.

The differences in meaning between light verbs such as take in (4.17) and give in (4.18)

are explained in terms of different force-dynamics patterns. The overall flow in the

events described by the verbs is differently oriented in the two examples.

(4.17) Take a { sniff / taste } of the sauce, will you?

(4.18) Give the sauce a { sniff / taste }, will you?

In (4.17) it is the opinion of the addressee that is asked for, so that the energy is directed

towards the agent. This orientation corresponds to the force-dynamic pattern of the verb

take, which is a self-oriented activity. The question in (4.18) is about the sauce. One

wants to know whether it had spoiled. This direction corresponds to the pattern of the

verb give, which is a directed activity, oriented outwards with respect to the agent of

the event.

The account of Brugman (2001) provides a general framework for discussing the meaning

of light verbs. However, it does not relate the identified components of meaning with

139


the discussion concerning the degree of lightness of the verbs and the variety of light

verb constructions.

Kearns (2002) proposes a set of formal syntactic tests to distinguish between the con-

structions with “lighter” verbs and the constructions with “heavier” verbs. The former

group is called true light verb constructions and the latter group is called constructions

with vague action verbs. True light verb constructions are identified as special syntactic

forms, while the constructions with vague action verbs are regarded as regular phrases.

(4.19) a. The inspection was made by the man on the right.

b. * A groan was given by the man on the right.

(4.20) a. Which inspection did John make?

b. * Which groan did John give?

(4.21) a. I made an inspection and then Bill made one too.

b. * I gave the soup a heat and then Bill gave it one too.

The formal distinction between true light verbs and vague action verbs is illustrated in

(4.19-4.21), where the expression make an inspection represents constructions with vague

action verbs, and the expression give a groan represents true light verb constructions.

The examples show that the complement of a true light verb cannot be moved or omitted

in regular syntactic transformations. While the passive form of the expression make an

inspection (4.19a) is grammatical, the passive form of the expression give a groan (4.19b)

is not grammatical. The same asymmetry holds for the WH-question transformation in

(4.20) and for the co-ordination transformation in (4.21).

Kearns (2002)’s analysis points to some observable indicators on the basis of which

various light verb constructions can be differentiated and classified. However, it does

not relate the observed behaviour with the meaning of the verbs, regarding true light

verbs as semantically empty.

The empirical case study presented in this chapter addresses both issues discussed in

the literature: the components of meaning of light verbs discussed by Brugman (2001)

140

4.3. Experiments

and the degree of compositionality of light verb constructions discussed in the other

presented accounts. Following Grimshaw and Mester (1988) and Kearns (2002), we

distinguish between two kinds of constructions. We use Kearns (2002)’s terminology

referring to the more idiomatic constructions as true light verb constructions and to

the less idiomatic constructions as constructions with vague action verbs. We follow

Brugman (2001) in using force-dynamic schemata for describing the meaning of light

verbs. In the experiments presented in the following section, we examine the relationship

between the meaning of light verbs and their cross-linguistic syntactic behaviour.

An in-depth empirical study of light verb constructions in the specific context of par-

allel corpora and alignment can lead to new generalisations concerning the correlation

of their linguistic and statistical properties. On the one hand, the statistical large-scale

analysis of the behaviour of these constructions in a general cross-linguistic word align-

ment process provides novel linguistic information, which enlarges the empirical basis

for the analysis of these constructions, and complements the traditional grammaticality

judgments. On the other hand, the linguistically fine-grained analysis of the statistical

behaviour of these constructions provides linguistically-informed performance and error

analyses that can be used to improve systems for automatic word alignment.

4.3. Experiments

The purpose of our study is to examine the translation equivalents of a range of English

light verb constructions and the effect that lexical properties of light verbs have on the

cross-linguistic variation. We take as a starting point the observation that the cross-

linguistic distribution of light verb constructions depends on their structural properties,

as shown in (4.1-4.2), repeated here as (4.22-4.23).

(4.22) a. Mary [had a laugh]. (English)

b. Maria [lachte]. (German)

c. Marija se [na-smejala]. (Serbian)

(4.23) a. Mary [gave a talk]. (English)

141


b. Maria [hielt einen Vortrag]. (German)

c. Marija [je o-drzala predavanje]. (Serbian)

Recall that English light verb constructions are paraphrases of verbs. The expressions

had a laugh in (4.22a) and gave a talk in (4.23) can be replaced by the corresponding

verbs laughed and talked respectively without changing the meaning of the sentences.

(Obtaining natural sentences with the verbs instead of the constructions would require

adding some modifiers, but this does not influence their semantic equivalence.) The

corresponding cross-linguistic realisations of the constructions illustrated in (4.22-4.23)

can be either single verbs or constructions. The cross-linguistic variation can therefore

be seen as an extension of the within-language variation. We analyse the cross-linguistic

frequency distribution of the two alternants as an observable indicator of the lexical

properties of the constructions which spread across languages.

We explore the potential relationship between the meaning of light verbs and the cross-

linguistic realisations of light verb constructions by examining a sample of constructions

formed with two light verbs widely discussed in the literature. We select the verb take as

a representative of self-oriented force dynamic schemata (following Brugman (2001), as

discussed in Section 4.2.2). We select the verb make as a representative of directed force

dynamic schemata, similar in this respect to give, as analysed by Brugman (2001). The

reason for studying the verb make instead of give is to keep the number of arguments

constant across the verbs (give takes three arguments, while take takes two), excluding

this factor as a possible source of variation.

To compare the realisations of light verbs with realisations of regular lexical verbs,

we compose a set of verbs which are “heavy” lexical entries comparable in meaning

with the verb make. The set consists of the following verbs: create, produce, draw,

fix, (re)construct, (re)build, establish. It is obtained from WordNet (Fellbaum 1998),

which is a widely cited lexical resource specifying lexical relationships between words.

Including several representatives of regular verbs is necessary to deal with the differences

in frequency. Since the two light verbs are much more frequent than any of the regular

verbs, comparable samples cannot be drawn from corpora of the same size. For example,

in the same portion of a corpus which contains fifty occurrences of the verb make, one

can expect less than ten occurrences of the verb create. To obtain comparable samples,

142

4.3. Experiments

Figure 4.2.: Constructions with vague action verbs

we sum up the numbers of occurrences of all regular verbs regarding them as a single

regular verb during the analysis.

Our samples consist of instances of English light verbs and their German equivalents

automatically extracted from a word-aligned parallel corpus. We use this language pair

as a sample of many possible language pairs. In principle, the same analysis can be

performed for any given language pair.

We identify two aspects of the alignment of these constructions as the relevant objects of

study. First, we quantify the amount and nature of correct word alignments for light verb

constructions compared to regular verbs, as determined by human inspection. Given the

cross-linguistic variation between English, German, and Serbian, described in (4.1-4.2),

it can be expected that English light verb constructions will be aligned with a single word

more often than constructions headed by a regular verb. Assuming that the properties of

the heading light verbs do influence semantic compositionality of the constructions, it can

also be expected that light verb constructions headed by different verbs will be differently

aligned to the translations in other languages. Different patterns of alignment would thus

indicate different types of constructions. Second, we evaluate the quality of automatic

word alignments of light verb constructions. Translations that deviate from one-to-one

word alignments, as it is the case with light verb constructions, are hard to handle in

the current approaches to automatic word alignment (see Section 3.2.1 in Chapter 3).

Because of the cross-linguistic variation illustrated in (4.1-4.2), light verb constructions

can be expected to pose a problem for automatic word alignment. Specifically, we

expect lower overall quality of word alignment in the sentences containing light verb

constructions than in the sentences that contain corresponding regular constructions.

143


Figure 4.3.: True light verb constructions

4.3.1. Experiment 1: manual alignment of light verb constructions

in a parallel corpus

In the first experiment, we address the relationship between two distinctions pointed

out in the theoretical accounts of light verb constructions: a) the distinction between

self-oriented vs. directed dynamics in the meaning of light verbs and b) the distinc-

tion between idiomatic true light verb constructions vs. regular-like constructions with

vague action verbs. The experiment consists of manual word alignment and a statistical

analysis of a random sample of three kinds of constructions: constructions with the verb

take, constructions with the verb make, and regular constructions.

We test the following hypotheses:

1. Light verb constructions in English are aligned with a single word in German more

often than constructions headed by a regular verb.

2. True light verb constructions in English are aligned with a single word in German

more often than constructions with vague action verbs.

3. The degree of compositionality of light verb constructions depends on the force

dynamic schemata represented in the meaning of light verbs.

We assume that the lack of cross-linguistic parallelism indicates idiosyncratic structures.

In the case of light verb constructions, we assume that the one-to-two word alignment

illustrated in Figure 4.3 indicates idiomatic true light verb constructions, while the one-

to-one word alignment illustrated in Figure 4.2 indicates more regular constructions with

144

4.3. Experiments

vague action verbs. We assume that both types have some semantic content, but that

this content is richer in the latter group than in the former.

Materials and methods

We analyse three samples of the constructions, one for each of the types defined by the

heading verb. Each sample contains 100 instances randomly selected from a parallel

corpus. Only the constructions where the complement is the direct object were included

in the analysis. This means that constructions such as take something into consideration

are not included. The only exception to this were the instances of the construction take

something into account. This construction was included because it is used as a variation

of take account of something with the same translations to German. All the extracted

instances are listed in Appendix A.

Corpus. The instances of the phrases were taken from the English-German portion

of the Europarl corpus (Koehn 2005). The texts in Europarl are collected from the

website of the European Parliament. They are automatically segmented into sentences

and aligned at the level of sentence. The version of the corpus which we use contains

about 30 million words (1 million sentences) of each of the 11 formerly official languages

of the European Union: Danish (da), German (de), Greek (el), English (en), Spanish

(es), Finnish (fin), French (fr), Italian (it), Dutch (nl), Portuguese (pt), and Swedish

(sv). Most of the possible language pairs are not direct translations of each other,

since for each text, there is one source language and the others are translations. Some

translations are also mediated by a third language. All the instances analysed in this

study are extracted from the portion of the corpus which contains the proceedings of the

sessions held in 1999. The selected portion of the corpus is parsed using a constituent

parser (Titov and Henderson 2007).

Sampling. Instances of light verb constructions are sampled in two steps. First, a

sample of random 1000 bi-sentences is extracted using a sampler based on computer-

generated random numbers (Algorithm 1). Each sentence is selected only once (sampling

without replacement). All verb phrases headed by the verbs take and make, as well as the

six regular verbs in the randomly selected 1000 bi-sentences are extracted automatically.

145


1. Extract verb-noun pairs from the randomly selected automat-ically parsed sentences; Tgrep2 query:

’VP < ‘/^VB/ & <-1 (/^NP/ < (‘/^NN/ !. /^NN/))’

3. Select light verb construction candidates:a. the pairs which contain the verb take and a deverbalnominal complement listed in NOMLEXb. the pairs which contain the verb make and a deverbalnominal complement listed in NOMLEX.

4. Select the pairs which contain one of the regular verbs.

Figure 4.4.: Extracting verb-noun combinations

The extraction is performed in several steps, as summarised in Figure 4.4. We first

extract all the verb-noun pairs using Tgrep2, a specialised search engine for parsed

corpora (Rohde 2004). We formulate the query shown in Figure 4.4 to extract all the

verbs which head a verb phrase containing a nominal complement together with the head

of the complement. The noun which is immediately dominated by the noun phrase and

which is not followed by another noun is considered the head of the noun phrase. The

extracted verb noun pairs are then compared with the list of deverbal nominals in the

NOMLEX database (Macleod et al. 1998) to select the pairs which consist of one of the

light verbs and a nominalisation. The selected pairs are then manually examined and

uses which are not light are removed from the list. Regular constructions are extracted

from the verb-noun pairs by comparing the heading verb with our predefined sample of

regular verbs.

After assessing the frequency of the selected constructions in the initial random sample

of 1000 sentences, we assess that the number of sentences needed to extract 100 instances

of each of the three types of constructions is 6000. In the second step, we add 5000 ran-

domly selected bi-sentences to the initial sample using the same sampler and repeat the

extraction procedure. The final sample which was analysed in the experiment consists

of the first 100 occurrences of each construction type in the sample of 6000 randomly

selected bi-sentences.

146

4.3. Experiments

Algorithm 1: Selecting a random sample of bi-sentencesInput : Aligned corpus of bi-sentences S,

each sentence s ∈ S is assigned a unique number n

Output : A random sample of K bi-sentences

for i = 1 to i = K do

generate a random number r in the range from 1 to S;

for j = 1 to j = S do

if r(i) == n(s) then

select s;

remove s from S;

break;

end

end

end

Feature representation. The constructions are represented as ordered pairs of words

V + N, where the first word is the verb that heads the construction and the second is

the noun that heads the verb’s complement. For a word pair in English, we identify

the corresponding word or word pair in German which is its actual translation in the

parallel corpus. If either the English or the German verb form included auxiliary verbs

or modals, these were not considered. Only the lexical part of the forms were regarded

as word translations.

(4.24) Erhe

hatAUX

einena

Vorschlagproposal

gemacht.made

He made a proposal.

(4.25) English instance: made + proposal

German alignment: Vorschlag + gemacht (note that hat is left out)

Type of mapping: 2-2

We then determine the type of mapping between the translations. If the German

translation of an English word pair includes two words too (e.g. take+decision ↔Beschluss+fassen), this was marked as the “2-2” type. If German translation is a single

word, the mapping was marked with “2-1”. This type of alignment is further divided

147


EnglishLVC take LVC make Regular

Ger

man

tran

slat

ion 2 → 2 57 50 94

2 → 1N 8 18 22 → 1V 30 28 22 → 0 5 4 2

Total 100 100 100

Table 4.1.: Types of mapping between English constructions and their translation equiv-alents in German.

into “2-1N” and “2-1V”. In the first subtype, the English construction corresponds to a

German noun (e.g. initiative+taken ↔ Initiative). In the second subtype, the English

construction corresponds to a German verb (e.g. take+look↔ anschauen). In the cases

where a translation shift occurs so that no translation can be found, the mapping is

marked with “2-0”.

For example, a record of an occurrence of the English construction “make + proposal”

extracted from the bi-sentence in (4.24) would contain the information given in (4.25).

For more examples, see Appendix A.

Results and discussion

We summarise the collected counts in a contingency table and compare the observed

distributions with the distributions which are expected under the hypothesis that the

type of the construction does not influence the variation.

χ2 =∑ (E −O)2

E(4.26)

To asses whether the difference between the observed and the expected distributions

is statistically significant, we use the χ2-test, which is calculated using the equation in

(4.26), where O stands for observed counts, E for expected counts.

148

4.3. Experiments

Table 4.1 shows how many times each of the four types of mapping (2-2; 2-1N; 2-1V;

2-0) between English constructions and their German translation equivalents occurs in

the sample.

We can see that the three types of constructions tend to be mapped to their German

equivalents in different ways. First, both types of light verb constructions are mapped to

a single German word much more often than the regular constructions (38 instances of

light verb constructions with take and 46 instances of light verb constructions with make

vs. only 4 instances of regular constructions.). This difference is statistically significant

(χ2 = 56.89, p < 0.01). Confirming our initial hypothesis No. 1, this result suggests

that the difference between fully compositional phrases and light verb constructions

in English can be described in terms of the amount of the “2-1” mapping to German

translation equivalents.

The number of “2-1” mappings is not significantly different between light verb construc-

tions headed by take and those headed by make (χ2 = 4.54, p < 0.90). However, an

asymmetry can be observed concerning the two subtypes of the “2-1” mapping. The

German equivalent of an English construction is more often a verb if the construction

is headed by the verb take (in 30 occurrences, that is 79% of the 2-1 cases) than if the

construction is headed by the verb make (28 occurrences, 61% cases). This difference is

statistically significant (χ2 = 3.90, p < 0.05).

When the German translation equivalent for an English construction is a verb, the mean-

ing of both components of the English construction are included in the corresponding

German verb, the verbal category of the light verb and the lexical content of the nominal

complement. These instances are less compositional, more specific and idiomatic (e.g.

take+care ↔ kummern, take+notice ↔ berucksichtigen).

On the other hand, English constructions that correspond to a German noun are more

compositional, less idiomatic and closer to the regular verb usages (e.g. make+proposal

↔ Vorschlag, make+changes ↔ Korrekturen). The noun that is regarded as their Ger-

man translation equivalent is, in fact, the equivalent of the nominal part of the con-

struction, while the verbal part is simply omitted. This result suggests that English

light verb constructions with take are less compositional than the light verb construc-

tions with make.

149


This result does not confirm the hypothesis No. 2, but it does confirm the hypothesis

No. 3. Although the number of “2-1” mappings is not different between the two light

verbs, two kinds of these mappings can be distinguished. The statistically significant

difference in the mappings suggests that the degree of compositionality of light verb

constructions depends on the force dynamic schemata represented in the meaning of

light verbs. The agent-oriented dynamics of the verb take gives rise to more divergent

cross-linguistic mappings than the directed dynamics of the verb take.

4.3.2. Experiment 2: Automatic alignment of light verb

constructions in a parallel corpus

In the second experiment, we address the relationship between the degree of composi-

tionality of light verb constructions and the quality of automatic word-alignment. On

the basis of the results of the first experiment and of the assumption that divergent

alignments are generally more difficult for an automatic aligner than the one-to-one

alignments, we expect the quality of automatic alignment to depend on the heading

verb. In particular, we test the following hypotheses:

1. The quality of word alignment in the sentences containing light verb constructions

is lower than in the sentences that contain corresponding regular constructions.

2. The quality of word alignment in the sentences containing light verb constructions

headed by take is lower than in the sentences that contain light verb constructions

headed by make.


Corpus and sampling. The same sample of sentences as in the first experiment

is analysed. Before sampling, the corpus was word-aligned in both directions using

GIZA++ (Och and Ney 2003). As discussed in Section 3.2.1 in Chapter 3, the formal

definition of alignment used by this system excludes the possibility of aligning multiple

words in one language to multiple words in the other language, which is an option needed

150

4.3. Experiments

for representing alignment of non-compositional constructions. However, it does provide

the possibility of aligning multiple words in one to a single word in the other language,

which is the option needed to account for some of the described divergences between

English and German, such as the mappings shown in Figure 4.3. Such alignment is

possible in the setting where English is the target and German is the source language,

since in this case, both English words, the light verb and its complement can be aligned

with one German word. By contrast, if German is the target language, its single verb

that can be the translation for the English construction cannot be aligned with both

English words, but only with one of them. The direction of alignment can influence the

quality of automatic alignment, since the probability of alignment can only be calculated

for the cases that can be represented by the formal definition of alignment. The definition

of alignment implies that all the words in the target language sentence are necessarily

aligned, while some of the source sentence words can be left unaligned. This is another

reason why the quality of alignment can depend both on the type of the constructions

and on the direction of alignment.

Taking only the intersection of the alignments of both directions as the final automatic

alignment is a common practice. Its advantage is that it provides almost only good

alignments (precision 98.6% as evaluated by Pado (2007) and Och and Ney (2003)),

which can be very useful for some tasks. However, it has two disadvantages. First,

many words are left unaligned (recall only 52.9%). Second, it excludes the possibility

of many-to-one word alignment that is allowed by the alignment model itself and that

could be useful in aligning segments such as constructions with true light verbs. We

therefore do not use the intersection alignment, but rather analyse both directions.

(4.27) Target language German

EN: He made a proposal.

DE: Er(1) hat(1) einen(3) Vorschlag(4) gemacht(3).

Target language English

DE: Er hat einen Vorschlag gemacht.

EN: He(1) made(5) a(3) proposal(4).

(4.28) Automatic alignment, target German, noun: good, verb: no align

Automatic alignment, target English, noun: good, verb: good

151


Figure 4.5.: The difference in automatic alignment depending on the direction.

Alignment categories. We examine the output of the automatic aligner for the sample

of 300 instances described in Section 4.3.1 comparing it with the manual alignment

obtained in the first experiment. We collect the information on automatic alignment for

each element of the English word pair for both alignment directions. The alignment was

assessed as “good” if the construction or the individual word is aligned with its actual

translation, as “bad” if the construction or the word is aligned with some other word, and

as “no align” if no alignment is found. For example, the automatically aligned sentences

in (4.27) would be recorded as in (4.28) (The numbers in the brackets represent the

positions of the aligned words). More examples can be found in Appendix A. Note that

the “no align” label can only occur in the setting were English is the source language,

since all the words in the sentence have to be aligned in the case when English is the

target language.


We evaluate the quality of automatic alignment comparing the alignment of the three

types of constructions and taking into account the effects of the direction of alignment.

152

4.3. Experiments

Target DE Target EN

LVCs withtake

Both EN words 5 57EN noun 63 79EN verb 6 57

LVCs withmake


Regularconstruction


Table 4.2.: Well-aligned instances of light verb constructions with take, with make, andwith regular constructions (out of 100), produced by an automatic alignment,in both alignment directions (target is indicated).

As in the first experiment, the statistical significance of the observed differences in

frequency distributions is assessed using the χ2-test.

Table 4.2 shows how the quality of automatic alignment varies depending on the type of

construction, but also on the direction of alignment (see also Figure 4.5). Both words are

well aligned in light verb constructions with take in 57 cases and with make in 40 cases

if the target languages is English, which is comparable with regular constructions (42

cases). However, if the target language is German, both types of light verb constructions

are aligned well (both words) in only 5 cases, while regular constructions are well aligned

in 26 cases.

The effect of the direction of alignment is expected in light verb constructions given the

underlying formal definition of alignment which does not allow multiple English words

to be aligned with a single German word when German is the target language. However,

the fact that the alignment of regular phrases is degraded in this direction too shows that

the alignment of light verb constructions influences other alignments. The difference in

the amount of correct alignments in two directions also shows the amount of the correct

alignments which remain out of the intersection alignment.

Looking into the alignment of the elements of the constructions (verbs and nouns) sepa-

153


Frequency take LVC make LVC RegularLow 12 25 62High 76 40 8

Table 4.3.: The three types of constructions partitioned by the frequency of the comple-ments in the sample.

rately, we can notice that nouns are generally better aligned than verbs for all the three

types of constructions, and in both directions. However, this difference is not the same

in all the cases. The difference in the quality of alignment of nouns and verbs is the

same in both alignment directions for regular constructions, but it is more pronounced

in light verb constructions if German is the target. On the other hand, if English is the

target, the difference is smaller in light verb construction than in regular phrases. These

findings suggest that the direction of alignment influences more the alignment of verbs

than the alignment of nouns in general. This influence is much stronger in light verb

constructions than in regular constructions.

Given the shown effects of the direction of alignment, we focus only on the direction

which allows for better alignments in all three groups (with English as the target lan-

guage) and perform statistical tests only for this direction. The difference between

alignments of both members of the three types of constructions (both EN words in Ta-

ble 4.2) is statistically significant (χ2=6,97, p < 0.05). However, this does not confirm

the initial hypothesis No. 1 that the quality of alignment of light verb constructions is

lower than the quality of alignment of regular constructions. The quality of alignment in

light verb constructions is, in fact, better than in regular constructions. The difference

in the quality of automatic alignment between the two kinds of light verb constructions

is also statistically significant (χ2=5.74, p < 0.05), but the difference is again opposite

to the hypothesis No. 2: constructions with take are better aligned than constructions

with make. On the other hand, there is no significant difference between constructions

with make and regular constructions. These results suggest that the type of construction

which is the least compositional and the most idiomatic of the three is best aligned if

the direction of alignment suits its properties.

154

4.3. Experiments

Figure 4.6.: The distribution of nominal complements in constructions with take. In 12out of 100 instance the complement is headed by a low-frequency noun (lowfrequency = 1 occurrence in the sample). There are 76 instances where thecomplement is headed by a high frequency noun: 5 (one noun with frequency5) + 7 (one noun with frequency 7) + 27 (three nouns with frequency 9) +17 (one noun with frequency 17) + 20 (one noun with frequency 20).

Figure 4.7.: The distribution of nominal complements in constructions with make. In25 out of 100 instance the complement is headed by a low-frequency noun(low frequency = 1 occurrence in the sample). There are 40 instances wherethe complement is headed by a high frequency noun: 15 (three nouns withfrequency 5) + 7 (one noun with frequency 7) + 8 (one noun with frequency8) + 10 (one noun with frequency 10).

155


Figure 4.8.: The distribution of nominal complements in regular constructions. In 62out of 100 instance the complement is headed by a low-frequency noun (lowfrequency = 1 occurrence in the sample. There are 8 instances where thecomplement is headed by a high frequency noun: one noun with frequency8.

Since the quality of alignment of the three types of constructions proved different from

what was expected in the case where English was the target language, we examine further

the automatic alignment in this direction. In particular, we investigate the influence of

the frequency distribution of the elements of light verb constructions on the quality of

alignment. This approach is based on the fact that the elements of idiomatic expressions

tend to occur more jointly than separately (Church and Hanks 1990). As discussed in

Section 3.2.1 in Chapter 3, the co-occurrence frequency is important for calculating

word-alignment, which is the factor that could have influenced the results. Since the

verb is a constant element within the three studied groups, we analyse the distribution

of the nominal complements.

The frequency of the nouns is defined as the number of occurrences in the sample. It

ranges from 1 to 20 occurrences in the sample of 100 instances. The instances of the

constructions are divided into three frequency ranges: instances containing nouns with

1 occurrence are regarded as low frequency items; those containing nouns that occurred

156

4.3. Experiments

Figure 4.9.: The difference in automatic alignment depending on the complement fre-quency. English is the target language.

5 and more times in the sample are regarded as high frequency items; nouns occurring

2, 3, and 4 times are regarded as medium frequency items. Only low and high frequency

items were considered in this analysis.

Table 4.3 shows the number of instances belonging to different frequency ranges. It

can be noted that light verb constructions with take exhibit a small number of low

frequency nouns (see also Figure 4.6). The number of low frequency nouns increases

in the constructions with make (25/100, see also Figure 4.7), and it is much bigger

in regular constructions (62/100, see also Figure 4.8). The opposite is true for high

frequency nouns (LVCs with take: 76/100, with make: 40/100, regular: 8/100). Such

distribution of low/high frequency items reflects different collocational properties of the

constructions. In the most idiomatic constructions (with take), lexical selection is rather

limited, which results in little variation. Verbs in regular constructions select for a wide

range of different complements with little reoccurrence. Constructions with make can

be placed between these two types.

Different trends in the quality of automatic alignment can be identified for the three

157


Well alignedFreq take lvc make lvc Regular

C % C % C %Low Both 4 33 8 32 21 34Freq N 8 66 8 32 47 75

V 4 33 12 48 53 85High Both 47 62 18 51 4 50Freq N 64 84 27 77 8 100

V 58 76 18 51 4 50

Table 4.4.: Counts and percentages of well-aligned instances of the three types of con-structions in relation with the frequency of the complements in the sample.The percentages represent the number of well-aligned instances out of theoverall number of instances within one frequency range. English is the targetlanguage.

types of constructions depending on the frequency range of the complement in the con-

structions, as shown in Table 4.4 and Figure 4.9.

First, the quality of alignment of both components of the constructions jointly is the same

for all the three types of constructions in low frequency items ( there is no statistically

significant difference between 33% well-aligned instances of light verb constructions with

take, 32% of light verb constructions with make, and 34% of regular constructions. The

alignment in this category is improved in high frequency items in all the three types,

compared to low frequency. The improvement is statistically significant (χ2 = 16.24 p <

0.01) Note that the high frequency regular items are represented with only 8 instances,

which is why the trends might not be clear enough for this subtype.

The analysis of the influence of frequency of verbs’ complements on the quality of auto-

matic alignment shows that frequency of words is more important for automatic align-

ment than the structural parallelism between languages. The alignment is significantly

better for high frequency combinations in all three types. Contrary to our hypothesis,

the idiomatic nature of light verb constructions with take does not pose a problem for

an automatic aligner due to the fact that a big proportion of instances of these construc-

tions belongs to the high frequency category. The quality of alignment in constructions

158

4.4. General discussion

with take is better than the quality in the other two types due to the difference in the

distribution of high frequency items. As it can been seen in Figures 4.6, 4.7, and 4.8,

the sample of constructions with take consists mostly of high frequency items. Low

frequency items, on the other hand, prevail in regular constructions, while constructions

with make are in between the two.


The results of our study confirm the hypotheses about the relationship between cross-

linguistic alignment of light verb constructions and the meaning of the heading light

verb tested in Experiment 1. On the other hand, the hypotheses about the relationship

between the type of light verb constructions and automatic word alignment in a parallel

corpus tested in Experiment 2 are not confirmed. However, the identified behaviour

of the automatic aligner with respect to light verb constructions provides additional

evidence for the distinctions confirmed in Experiment 1. In this section, we interpret

the results in light of the theoretical discussion concerning light verb constructions.

4.4.1. Two force dynamics schemata in light verbs

The main finding of the study is the fact that the constructions headed by light take

behave as idiomatic phrases more than the constructions headed by make. The differ-

ence between more idiomatic and less idiomatic light verb constructions has been widely

discussed in the literature, especially from the point of view of analysing the predicate-

argument structure of the constructions. The structure of true light verb constructions,

which are idiomatic and non-compositional, is argued to be similar to complex predicates,

while the structure of constructions with vague action verbs, which are less idiomatic

and compositional, is argued to be similar to regular phrases. Our study shows that

the idiomatic properties of light verb constructions can be related to the meaning of the

heading verbs. The self-oriented force dynamics in the meaning of light take results in

more compact cross-linguistic morphosyntactic realisations than the directed dynamics

of light make. Cross-linguistic equivalents of English light verb constructions with take

159


tend to be single verbs (which is a compact representation), while cross-linguistic equiv-

alents of English light verb constructions with make tend to stay constructions with two

main elements. This does not hold only for the language pair English-German, but also

for the pair English-Serbian, as discussed by Samardzic (2008).

The idiomatic nature of true light verb constructions (represented by the constructions

with take in our study) is additionally confirmed by the finding that these construc-

tions are better aligned automatically than regular constructions. This finding, which

is opposed to our hypotheses, is due to the same interaction between frequency and ir-

regularity which has been established in relation with different language processing and

acquisition phenomena. Idiosyncratic (irregular) elements of language are known to be

more frequent than regular unites. This is the case, for example, with English irregular

verbs, which are, on average, more frequent than regular verbs. In the case of light

verb constructions in our study, the idiosyncratic units are the constructions with take

which are idiomatic with high co-occurrence of the two elements (the heading verb and

the nominal complement). The constructions with make, which represent constructions

with vague action verbs in our study, can be positioned somewhere between irregular and

regular items. This additionally confirms the claim that these two types of constructions

differ in the level of semantic compositionality.

Our analysis of corpus data has shown that there is a clear difference between regular

phrases and light verb constructions (including the constructions with make) in the

way they are cross-linguistically mapped in a parallel corpus. Regular constructions

are mapped word-by-word, with the English verb being mapped to the German verb,

and the English noun to the German noun. A closer look into the only 4 examples

where regular constructions were mapped as “2-1” shows that this mapping is not due

to the “lightness” of the verb. In two of these cases, it is the content of the verb that

is translated, not that of the noun (produce+goods ↔ Produktion; establishes+rights

↔ legt). This never happens in light verb constructions. On the other hand, light

verb constructions are much more often translated with a single German word. In both

subtypes of the “2-1” mapping of light verb constructions, it is the content of the nominal

complement that is translated, not that of the verb. The noun is either transformed into

a verb (take+look↔ anschauen) or it is translated directly with the verb being omitted

(take+initiative ↔ Initiative).

160


The frequency distribution observed in out data represents a new piece of empirical ev-

idence of the distinctions made. The observable differences in cross-linguistic alignment

are especially useful for distinguishing between regular constructions and constructions

with vague action verbs (represented in our sample by the constructions with make). It

has been shown in other studies that true light verb constructions have characteristic

syntactic behaviour. Constructions with vague action verbs, however, can not be distin-

guished using the same tests, while they are clearly distinguished on the basis of their

cross-linguistic mappings.

4.4.2. Relevance of the findings to natural language processing

The findings of our study show that the interactions between automatic alignment and

types of constructions is actually more complicated than the simple hypotheses which we

initially formulated. To summarise, we find, first, better alignment of regular construc-

tions compared to light verb constructions only if the target language is German; second,

overall, alignment if English is target is better than if German is target; and third, we

found a clear frequency by construction interaction in the quality of alignment.

The quality of automatic alignment of both regular constructions and light verb con-

structions interacts with the direction of alignment. First, the alignment is considerably

better if the target language is English than if it is German, which confirms the findings

of Och and Ney (2003). Second, the expected difference in the quality of alignment

between regular constructions and light verb constructions has only been found in the

direction of alignment with German as the target language, that is where the “2-1”

mapping is excluded. However, the overall quality of alignment in this direction is lower

than in the other.

This result could be expected, given the general morphological properties of the two

languages, as well as the formalisation of the notion of word alignment used in the system

for automatic alignment. According to this definition, multiple words in the target

language sentence can be aligned with a single word in the source language sentence, but

not the other way around. Since English is a morphologically more analytical language

161


than German, multiple English words often need to be aligned with a single German

word (a situation allowed if English is the target but not if German is the target).

The phrases in (4.29) illustrate the two most common cases of such alignments. First,

English tends to use functional words (the preposition of in (3a)), where German applies

inflection (genitive suffixes on the article des and on the noun Bananensektors in (3b).

Second, compounds are regarded as multiple words in English (banana sector), while

they are single words in German (Bananensektors). This asymmetry explains both the

fact that automatic alignment of all the three types of constructions is better when the

target language is English and that the alignment of light verb constructions is worse

than the alignment of regular phrases when it is forced to be expressed as one-to-one

mapping, which occurs when German is the alignment target.

(4.29) a. the infrastructure of the banana sector

b. die Infrastruktur des Bananensektors

Practically, all these factors need to be taken into consideration in deciding which version

of alignment should be taken, be it for evaluation or for application in other tasks such as

automatic translation or annotation projection. The intersection of the two directions

has been proved to provide most reliable automatic alignment Pado (2007); Och and

Ney (2003). However, it excludes, by definition, all the cases of potentially useful good

alignments that are only possible in one direction of alignment.

4.5. Related work

Corpus-based approaches to light verb constructions belong to the very developed do-

main of collocation extraction. General methods developed for automatic identification

of collations in texts based on various measures of association between words can also be

applied to light verb constructions. However, light verb constructions differ from other

types of collocations in that they are partially compositional and relatively productive,

which calls for a special treatment.

162

4.5. Related work

The methods which combine syntactic parsing with standard measures of association

between words prove to be especially well adapted for automatic identification of light

verb constructions (Seretan 2011). Identifying the association between syntactic con-

stituents rather than between the words in a context window allows identifying light

verb constructions as collocations despite the variation in their realisations due to their

partially compositional meaning.

Grefenstette and Teufel (1995) present a method for automatic identification of an appro-

priate light verb for a derived nominal on the basis of corpus data. Making a difference

between the cases where the derived nominals can be ambiguous between more verb-like

(e. g. make a proposal) and more noun-like (e.g. put the proposal in the drawer) uses,

Grefenstette and Teufel (1995) extract only those usages where the noun occurs in a

context similar to a typical context of the corresponding verb. The most frequent gov-

erning verbs for these noun occurrences are their light verbs. As noted by the authors,

this technique proves to be insufficient on its own for identifying light verbs. It does

not differentiate between the light verb and other frequent verbal collocates for a given

nominalisation (e. g. reject a proposal vs. make a proposal). But it can be used as

a step in automatic processing of corpora, since light verbs do occur in the lists of the

most frequent collocates.

The method for extracting verb-noun collocations proposed by Tapanainen et al. (1998)

is based on the assumption that collocations of the type verb-noun are asymmetric in

such a way that it is the object (i. e. the noun) that is more indicative of the construction

being a collocation. If a noun occurs as the object of only few verbs in a big corpus,

its usage is idiomatical. For example, the noun toll occurs mainly with the verb take.

It can be used with other verbs too (e. g. charge, collect), but not with many. The

measure proposed in the study, the distributed frequency of the object, is better suited

for extracting light verb constructions than some symmetric measures of association.

However, this approach does not provide a means to distinguish light verb constructions

from the other collocations of the same type.

Using the information from cross-linguistic word alignment for identifying collocations

is explored by Zarrieß and Kuhn (2009). The study shows that many-to-one automatic

word alignment in parallel corpora is a good indicator of reduced compositionality of ex-

163


pressions. Combined with syntactic parsing, this information can be used for automatic

identification of a range of collocation types, including light verb constructions.

Semantic characteristics of light verb constructions are studied in more detail by Fazly

(2007), who proposes a statistical measure that quantifies the degree of figurativeness of

the light verb in conjunction with a predicating noun. The degree of figurativeness of a

verb is regarded as the degree to which its meaning is different from its literal meaning

in a certain realisation. It is assumed that constructions of the type verb-noun can be

placed on a continuum of figurativeness of meaning, including literal combinations (e. g.

give a present), abstract combinations (e. g. give confidence), light verb constructions

(e. g. give a groan), and idiomatic expressions (e. g. give a whirl). More figurative

meanings of the verb are considered closer to true light verbs, while more literal meanings

are closer to vague action verbs (i. e. to the abstract combinations on the presented

continuum).

The measure of figurativeness is based on indicators of conventionalised use of the con-

structions: the more the two words occur together and the more they occur within a

particular syntactic pattern the more figurative the meaning of the verb. Thus, the

figurativeness score is composed of two measures of association: association of the two

words and association of the verb with a particular syntactic pattern.

The syntactic pattern that is expected for figurative combinations is defined in terms of

three formal properties associated with typical light verb constructions (see the examples

in (4.19-4.21) in Section 4.2.2): active voice, indefinite (or no) article, and singular form

of the noun. The association of a verb-noun combination with the expected syntactic

pattern is expressed as the difference between the association of the combination with

this pattern (positive association) and the association of the combination with any of

the patterns where any of the features has the opposite value (passive voice, definite

article, plural noun).

For a sample of expressions, the scores assigned by the measure of figurativeness are

compared with the ratings assigned by human judges. The results show that a measure

which includes linguistic information about expressions performs better in measuring

the degree of their figurativeness than a simple baseline measure of association between

the words in the expressions.

164

4.6. Summary of contributions

The work of Stevenson et al. (2004) deals with semantic constraints on light verb com-

plements. They focus on true light verb constructions trying to identify the classes of

complements that would be preferred by a given light verb. Light verb constructions are

first identified automatically and then the relations between light verbs and some classes

of complements are examined. Following the analysis of Wierzbicka (1982) (see Section

4.2.1), the nominal complements of light verbs are identified with their corresponding

verbs. With this, it was possible to use Levin’s lexical semantic classification of verbs

(see Section 2.2.1 in Chapter 2 for more details) to divide the complements into semantic

classes and to examine if certain light verbs prefer certain classes of complements.

The study shows that light verbs have some degree of systematic and predictable be-

haviour with respect to the class of their complement. For example, light give tend to

combine with deverbal nouns derived from the Sound Emission verbs, while light take

combines better with the nouns derived from the Motion (non-vehicle) verbs. As the

light verb construction score gets higher, the pattern gets clearer. It shows as well that

some of the verbs (e. g. give and take) behave in a more consistent way than others (e.

g. make).

The computational approaches presented in this section show that the compositionality

of the meaning of light verb constructions does not correspond directly to the strength of

the association of their components. Adding specific linguistic information improves the

correlation between human judgements and automatic rankings. The studies, however,

do not address the lexical properties of light verbs verbs as one of the potential causes

of the observed variation. Also, they do not address the patterns in cross-linguistic

variation which are potentially caused be caused by different degrees of compositionality

of light verb constructions. Our study focuses on these two issues.


In the study of light verb constructions, we have proposed using data automatically

extracted from parallel corpora to identify two kinds of meaning of light verbs. We have

shown that English light verb constructions headed by the verb take tend to be aligned

165


with a German single verb more than the constructions headed by the verb make. The

difference in the cross-linguistic mapping is predicted from the meaning of the verbs

described in terms of force dynamics: the self-oriented schemata of light take gives rise

to more compact cross-linguistic realisations than the directed schemata of the verb

make.

The difference in the force dynamics of the two verbs is related to the level of composi-

tionality of their corresponding light verb constructions. The constructions with take are

less compositional and more irregular than the constructions with make. The idiomatic

nature of light verb constructions represented in our study by the constructions with

take is additionally confirmed by the finding that these constructions are better auto-

matically aligned than the constructions represented with the verb make, as well as the

comparable regular constructions. Although this finding sounds surprising, it actually

follows from the interaction of frequency and regularity which plays an important role

in automatic word alignment.

166

5. Likelihood of external causation

and the cross-linguistic variation in

lexical causatives

5.1. Introduction

The causative(/inchoative) alternation has been recognised in the linguistic literature a

wide- spread linguistic phenomenon, attested in almost all languages (Schafer 2009).

This alternation involves verbs such as break in (5.1), which can be realised in a sentence

both as transitive (5.1a) and as intransitive (5.1b). Both realisations express the same

event, with the only difference being that the transitive version specifies the causer of

the event (Adam in (5.1a)), and the intransitive version does not. The transitive version

is thus termed causative and the intransitive anticausative. The verbs that participate

in this alternation are commonly referred to as lexical causatives.1

(5.1) a. Causative: Adam broke the laptop.

b. Anticausative: The laptop broke.

1The lexical causative alternation, which we address in this study, is to be distinguished from thesyntactic causative alternation illustrated in (i), which has been studied more extensively in thelinguistic literature, as a case of verb serialisation (Baker 1988; Williams 1997; Alsina 1997;Collins 1997; Aboh 2009).

(i.) a. Lexical causative: Adam broke the laptop.

b. Syntactic causative: Adam made the laptop break.

167

5. Likelihood of external causation and the cross-linguistic variation in lexical causatives

What makes this alternation an especially attractive topic for our research is the wide

range of cross-linguistic variation in surface forms of the clauses formed with the alter-

nating verbs. The causative alternation appears in different languages with a diversity

of lexical, morphological, and syntactic realisations which defies linguists’ attempts at

generalisation.

First of all, the variation is observed in the sets of alternating verbs. Most of the

alternating verbs are lexical counterparts across many languages. However, there are

still many verbs which alternate in some languages, while their lexical counterparts in

other languages do not. The verbs that do not alternate in some languages can be divided

into two groups: only intransitive and only transitive. Examples of only intransitive and

only transitive verbs in English are given in Table 5.1. As the examples, taken from

Alexiadou (2010), show, the English verbs arrive and appear do not alternate: their

transitive realisation (causative in Table 5.1) is not available in English. However, their

counterparts in Japanese, or Salish languages, for example, are found both as transitive

and intransitive, that is as alternating. Similarly, the verbs such as cut and kill are only

found as transitive in English, while their counterparts in Greek or Hindi, for example,

can alternate between intransitive and transitive use.

Languages also differ in the morphological realisation of the alternation. Some exam-

ples of morphological variation, taken from Haspelmath (1993), are given in Table 5.1.

In some languages, such as Russian, Mongolian, and Japanese, the alternation is mor-

phologically marked. The morpheme that marks the alternation can be found on the

intransitive form, while the corresponding transitive form is not marked (the case of Rus-

sian in Table 5.1). In other languages, such as Mongolian, the morpheme that marks the

alternation is found on the transitive version, while the intransitive version is unmarked.

There are also languages where both forms are attached a causative marker, one mark-

ing the transitive and the other the intransitive version, like in the Japanese example in

Table 5.1. English, on the other hand, is an example of a language where the alternation

is not marked at all. (Note that both forms of the verbs melt and gather in Table 5.1

are the same.) The different marking strategies illustrated in Table 5.1 represent only

the most common markings. Languages can use different options for different verbs. For

example, anticausative versions of some verbs are not marked in Russian. In principle,

any option can be found in any language, but with different probability.

168

5.1. Introduction

Availability :Causative Anticausative

arrive, appear +Japanese, +Salish,-English

+all languages

kill, cut +all languages +Greek, +Hindi,-English

Morphological marking :Causative Anticausative

Mongolian xajl-uul-ax xajl-ax’melt’ ’melt’

Russian rasplavit’ rasplavit’-sja’melt’ ’melt’

Japanese atum-eru atum-aru’gather’ ’gather’

Table 5.1.: Availability of the alternation (Alexiadou 2010) and morphological marking(Haspelmath 1993) in some examples of verbs and languages.

169


The variation in the availability of the alternation illustrated in Table 5.1 raises the ques-

tion of why some verbs do not alternate in some languages. Answering this question

can help understanding why alternating verbs do alternate. The variation in morpho-

logical marking is even more puzzling: Why is it that languages do not agree on which

version of the alternating verbs to mark? Also, what needs to be addressed is the inter-

action between the categories of variation: Is there a connection between the alternation

availability and morphological marking?

In this study, we address the issues raised by the causative alternation in a novel approach

which combines the knowledge about the use of verbs in a corpus with the knowledge

about the typological variation.2 We analyse within-language variation in realisations of

lexical causatives, as well as cross-linguistic variation in a parallel corpus with the aim of

identifying common properties of lexical causatives underlying the variation. Our anal-

ysis relates the variation observed in language corpora with the observed cross-linguistic

variation in a unified account of the lexical representation of lexical causatives. The

findings of our study are expected to extend the knowledge about the nature of the

causative alternation by taking into consideration much more data than the previous

accounts. On the other hand, they are also expected to be applicable in natural lan-

guage processing. Being able to predict the cross-linguistic transformations of phrases

involving lexical causatives based on their common lexical representation can be useful

for improving automatic alignment of phrase constituents, which is a necessary step

in machine translation and other tasks in cross-linguistic language processing. Having

both purposes in mind, we propose and account of lexical causatives suitable for machine

learning. We model all the the studied factors so that the values of the variables can be

learned automatically by observing the instances of the alternating verbs in a corpus.

The chapter is organised in the following way. We start by discussing the questions raised

by lexical causatives. In Section 5.2.1, we introduce the distinction between internally

and externally caused events in relation to the argument structure of lexical causatives

and to the causative alternation. In Section 5.2.2, we discuss cross-linguistic variation in

the causative alternation and the challenges that it poses for the account based on the

two-way distinction between internally and externally caused event. In Section 5.2.3,

we discuss a more elaborate typological approach to cross-linguistic variation in lexical

2Some pieces of the work presented in this chapter are published as Samardzic and Merlo (2012).

170

5.2. Theoretical accounts of lexical causatives

causatives which proposes an account of their meaning in terms of a scale, rather than

two or three classes. After defining the theoretical context of our study, we present our

experimental approach to the questions discussed in the literature. The study consists

of four experiments. The first two experiments (Sections 5.3.1 and 5.3.2) establish a

corpus-based measure which can be used to distinguish between lexical causatives with

different lexical representations. In the third experiment (Section 5.3.3), we examine

the influence of the meaning of lexical causatives on their cross-linguistic realisations. In

the fourth experiment (Section 5.3.4), we test a statistical model which classifies lexical

causatives based on their cross-linguistic realisations. In Section 5.4, we interpret the

results of our experiments in light of the theoretical discussion and also in relation to

more practical issues concerning natural language processing. We compare our study

with the related work in Section 5.5.


Among other issues raised by the causative alternation which have been discussed in

the linguistic literature, theoretical accounts have been proposed for the interrelated

questions which we address in our study:

1. What are the properties of alternating verbs that distinguish them from the verbs

that do not alternate?

2. Which one of the two realisations is the basic form and which one is the derivation?

3. What is the source of cross-linguistic variation in the sets of alternating verbs?

Most of the proposed accounts are focused on the specific properties of alternating verbs

(Question No. 1) and on the structural relationship between the two alternants (Ques-

tion No. 2). The issues in cross-linguistic variation are usually not directly addressed

except in typological studies. Our study addresses the issue of cross-linguistic variation

directly, but the findings are relevant for the other two questions too. In this section, we

present theoretical accounts of lexical causatives introducing the notions and distinctions

which are addressed in our study. We focus on the proposals and ideas concerning the

171


relationship between the meaning of lexical causatives and the variation in their mor-

phosyntactic realisations, especially cross-linguistic variation, leaving on the side the

accounts of the structure of clauses formed with these verbs.

The most apparent common property of the alternating verbs in different languages

is their meaning. Most of these verbs describe an event in which the state of one of

the participants (patient or theme) changes (Levin and Rappaport Hovav 1994). If a

verb describes some kind of change of state, it can be used both as causative and as

anticausative. This is the case illustrated in (5.1) repeated here as (5.2). In the causative

use (5.2a), the verb is transitive, the changing participant is its object, and the agent

is expressed as its subject. In the anticausative use, the verb is intransitive, with the

changing participant being expressed as its subject (5.2b).

(5.2) a. Adam broke the laptop.

b. The laptop broke.

If the change-of-state condition is not satisfied, the alternation does not take place. The

example in (5.3a) illustrates a case of an intransitive verb whose subject does not undergo

a change, which is why it cannot be used transitively, as shown in (5.3b). Similarly, the

object of the verb bought in (5.4a) is not interpreted as changing, so the verb cannot be

used intransitively (5.4b).

(5.3) a. The children played.

b. * The parents played the children.

(5.4) a. The parents bought the toys.

b. * The toys bought.

5.2.1. Externally and internally caused events

Taking other verbs into consideration, however, it becomes evident that the meaning

of change of state is neither necessary nor sufficient condition for the alternation to

take place. Verbs can alternate even if they do not describe a change-of-state event.

On the other hand, some verbs do describe a state-of-change event but they still do

172


not alternate. For example, verbs with the meaning of positioning such as hanging in

(5.5) do alternate although their meaning, at least in the anticausative version, does not

involve change of state. On the other hand, verbs such as transitive cut in (5.6) and

intransitive bloomed in (5.7) do not alternate although their meaning involves a change

of state, of bread in (5.6) and of flowers in (5.7).

(5.5) a. Their photo was hanging on the wall.

b. They were hanging their photo on the wall.

(5.6) a. The baker cut the bread.

b. * The bread cut.

(5.7) a. The flowers suddenly bloomed.

b. * The summer bloomed the flowers.

To deal with the issue of non-change-of-state verbs entering the alternation, Levin and

Rappaport Hovav (1994) introduce the notion of “externally” and “internally” caused

events. Externally caused events can be expressed as transitive forms, while internally

caused events cannot. On this account, verbs such as hanging in (5.5) can alternate even

though they do not describe a change of state event, because they mean something that

is externally caused. The hanging of the photo in (5.5a) is not initiated by the photo

itself, but by some other, external cause, which can then be expressed as the agent in a

transitive construction. The same distinction explains the ungrammaticality of (5.7b).

Since blooming is something that flowers do on their own, themselves, the verb bloom

does not specify an external causer which would be realised as its subject, which is why

this verb cannot occur in a transitive construction.

This distinction still does not account for all the problematic cases. It leaves without

an explanation the case of transitive verbs which describe a change of state, and which

are clearly externally caused, but which do not alternate such as cut in (5.6). To deal

with these cases, Levin and Rappaport Hovav (1994) introduce the notion of agentivity.

According to this explanation, the meaning of some verbs is such that specifying the

agent in the event described by the verb is obligatory, which is why they cannot occur

as intransitive. Levin and Rappaport Hovav (1994) argue that this happens with verbs

whose subject can only be the agent (and not the instrument, for example). Schafer

173


(2009) challenges this view showing that the alternation can be blocked in verbs with

different subjects. Such a verb is English destroy whose subject can be some natural

force, an abstract entity, or even an instrument, but which still does not alternate.

Haspelmath (1993) argues that it is the level of specificity of the verb that plays a role

in blocking the alternation. If a verb describes an event that is highly specified, such as

English decapitate, the presence of specific details in the interpretation of the meaning

of the verb can block the alternation.

5.2.2. Two or three classes of verb roots?

Since the discussed properties concern the meaning of verbs, one could expect that the

verbs which are translations of each other alternate in all languages. This is, however,

not always true. There are many verbs that do alternate in some languages, while

their counterparts in other languages do not. For example, in Greek and Hindi, the

counterparts of kill and destroy have intransitive versions (Alexiadou et al. 2006).

On the other hand, typically intransitive verbs of moving (run, swim, walk, fly) can

have transitive versions in English, which is not possible in French or German (Schafer

2009).

An explanation for these cases is proposed by Haspelmath (1993), who argues that a

possible cause of these differences is a slightly different meaning of the lexical coun-

terparts across languages. Russian myt’, for example, which does alternate, does not

mean exactly the same as English wash, which does not alternate. Haspelmath (1993),

however, does not propose a particular property which differs in the two verbs.

The question of cross-linguistic variation has received more attention in the work of

Alexiadou (2010) who examines a wide range of linguistic facts including the variation

in the availability of the alternation and in morphological marking. Alexiadou (2010)

argues that the account of the examined facts requires introducing one more class of

verbs3. In addition to the classes of externally caused and internally caused verbs,

proposed by Levin and Rappaport Hovav (1994), Alexiadou (2010) proposes a third

3More precisely, Alexiadou (2010) refers to verb roots rather than to verbs to emphasise that thediscussion concerns this particular level of the lexical representation.

174


Causative AnticausativeGreek: spao ’break’ spao ’break’

klino ’close’ klino ’close’aniyo ’open’ aniyo ’open’

Japanese: war-u ’break’ war-er-u ’break’Turkish : kapa ’close’ kapa-n ’close’

Table 5.2.: The examples of morphological marking of cause unspecified verbs discussedby Alexiadou (2010).

group of “unspecified roots”. The generalisations based on the proposed framework are

summarised in (5.8).

(5.8) a. Anticausative verbs that are characterised as internally caused and/or cause

unspecified are not morphologically marked, while those that are characterised as

externally caused are marked.

b. Cause unspecified verbs alternate in all languages, while internally caused and

externally caused verbs alternate only in languages that allow anticausative

morphological marking.

Although Alexiadou’s (2010) analysis relates the two important aspects of the crosslin-

guistic variation in an innovative way, it fails to explain the tendency observed in many

languages regarding the morphological marking of the anticausative variant. In partic-

ular, of all the examples mentioned in Alexiadou (2010) (see Table 5.2) as support to

(5.8a), only the Greek examples can be clearly classified. The other examples in Table

5.2 illustrate that verbs classified as prototypical cause-unspecified (Class I) (e.g., break,

open, close (Alexiadou et al. 2006; Alexiadou 2010)) tend to allow rather than disallow

morphological marking of their anticausative variant (compare the examples in Table

5.2).4

Looking up the verbs in other languages shows some limitations for the generalisation

in (5.8b) too. For example, the Serbian verb rasti ’grow’ would be classified as cause

4These verbs are mostly classified as Class II (externally caused) in the data overview, but they areclassified as Class I in the summary of the data.

175


unspecified according to Alexiadou’s 2010 criteria, implying that it is expected to alter-

nate in all languages. However, this verb does not alternate in Serbian; it exists only as

intransitive.

5.2.3. The scale of spontaneous occurrence

The approach proposed by Haspelmath (1993) does not address the syntactic aspects

of the variation, but it provides a better account of the data which pose a problem

for the generalisations proposed by Alexiadou (2010). Haspelmath (1993) analyses the

typology of morphological marking of the two realisations of alternating verbs across a

wade range of languages. Alternating verbs can be divided into several types according to

the morphological differences between the causative and the anticausative version of the

verb. The alternation can be considered morphologically directional if one form is derived

from the other.5 There are two directional types: causative, if the causative member of

the pair is marked, and anticausative, with morphological marking on the anticausative.

Morphologically non-directed alternations are equipollent, if both members of the pair

bear a morphological marker, suppletive, if two different verbs are used as alternants,

and labile if there is no difference in the form between the two verbs.

One language typically allows several types of marking, but it prefers one or two types.

For example, both English and German allow anticausative, equipollent, labile, and

suppletive alternations. There is a strong preference for labile pairs in English, while

German prefers anticausative and labile pairs (Haspelmath 1993).6

Despite the different morphological marking types of languages, a study of twenty-one

pairs of alternating verbs showed that certain alternating verbs tend to bear the same

5Certain authors (Alexiadou 2006a) argue against the direct derivation. Since precise account of thederivation of the structures is not relevant for our work, we maintain the morphological distinctionsdescribed in (Haspelmath 1993).

6The issue of whether these preferences can be related to some other properties of the languagesis still unresolved. The only correlation that could be observed is the fact that the anticausativemorphology is observed mostly in European languages, even if they are not closely genetically related.For example, Greek is as close to the other European languages as Hindi-Urdu. While Greek shows apreference for anticausative morphology, Hindi-Urdu prefers causative morphology. Languages thatprefer causative morphology are more spread, being located in almost all the continents, while thepreference for anticausative morphology is restricted mainly to Europe.

176


Languages (N) A C E L S A/Cboil 21 0.5 11.5 3 6 0 0.04freeze 21 2 12 3 4 0 0.17dry 20 3 10 4 3 0 0.30wake up 21 3 9 6 2 1 0.33go out / put out 21 3 7.5 5.5 3 2 0.41sink 21 4 9.5 5.5 1.5 0.5 0.42learn / teach 21 3.5 7.5 6 2 3 0.47melt 21 5 10.5 3 2.5 0 0.48stop 21 5.5 9 3.5 3 0 0.61turn 21 8 7.5 4 1.5 0 1.07dissolve 21 10.5 7.5 2 1 0 1.40burn 21 7 5 2 5 2 1.40destroy 20 8.5 5.5 5 1 0 1.50fill 21 8 5 5 3 0 1.60finish 21 7.5 4.5 5 4 0 1.67begin 19 5 3 3 8 0 1.67spread 21 11 6 3 1 0 1.83roll 21 8.5 4.5 5 3 0 1.89develop 21 10 5 5 1 0 2.00get lost / lose 21 11.5 4.5 4.5 0 0.5 2.56rise-raise 21 12 4.5 3.5 0 1 2.67improve 21 8.5 3 8 1.5 0 2.67rock 21 12 4 3.5 1.5 0 3.00connect 21 15 2.5 1.5 1 1 6.00change 21 11 1.5 4.5 4 0 7.33gather 21 15 2 3 1 0 7.50open 21 13 1.5 4 2.5 0 8.67break 21 12.5 1 2.5 2 0 12.50close 21 15.5 1 2.5 2 0 15.50split 20 11.5 0.5 5 3 0 23.00die / kill 21 0 3 1 1 16 —

Table 5.3.: Morphological marking across languages: A=anticausative, C=causative,E=equipollent, L=labile, S=suppletive

177


kind of marking across languages. Verbs such as lexical equivalents of English freeze, dry,

melt tend to be marked when used causatively in many different languages, while the

equivalents of English gather, open, break, close tend to be marked in their anticausative

uses. Table 5.3 shows the distribution of morphological marking for all the verbs included

in Haspelmath’s (1993) study. Note that the verbs are ranked according to the ratio

between anticausative and causative marking. The verbs with a low ratio are found on

the top of the table and those with a high ratio are in the bottom.

Assuming that the cross-linguistic distribution of the kinds of morphological marking

is a consequence of the way lexical items are used in language in general, Haspelmath

(1993) interprets these findings as pointing to a universal scale of increasing likelihood

of spontaneous occurrence. The verbs with a low A/C ratio describe events that are

likely to happen with no agent or external force involved. If the verb is used with an

expressed agent, the form of the verb contains a morphological marker in the majority

of languages. The verbs with a high A/C ratio typically specify an agent, and if the

agent is not specified, the verb tends to get some kind of morphological marking across

languages. In this interpretation, the cross-linguistic A/C ratio is an observable and

measurable indicator of a lexical property of verbs. It expresses the degree to which an

agent or an external cause is involved in the event described by the verb. A summary

of the notion of the scale of spontaneous occurrence is given in (5.9).

(5.9) The scale of spontaneous occurrence:

freeze > dry > melt > ..... > gather > open > break > close

low A/C (spontaneous) high A/C (non-spontaneous)

The notion of spontaneous occurrence can be related to the distinction between internally

and externally caused events argued for in the other analyses. Both notions concern the

same lexical property of verbs — the involvement of an agent in the event described by a

verb. The events that are placed on the spontaneous extreme of the scale would be those

that can be perceived as internally caused. The occurrence of an agent or an external

cause in these events is very unlikely. Since the externally caused events are considered

to give rise to the causative alternation, they would correspond to a wider portion of the

178

5.3. Experiments

scale of spontaneous occurrence, including not just the events on the non-spontaneous

extreme of the scale, but also those in the middle of the scale.

However, there are important theoretical and methodological differences between the

two conceptions. The qualitative notion of internal vs. external causation implies that

there are two kinds of events: those where the agent is present in the event and those

with no agent involved. Verbs describing internally caused event can only be used as

anticausative. A causative use of a verb describing an internally caused event is expected

to be ungrammatical (as in (5.7b)). The notion of scale of spontaneous occurrence

does not imply complete absence of the agent in any event. It does not predict the

ungrammaticality of uses such as (5.7b). What follows from this notion is that such uses

are possible, but very unlikely.

The difference between the two conceptions is even more important with respect to the

events perceived as externally caused. The qualitative analyses imply that all externally

caused events have the same status with respect to the causative alternation — the verbs

describing these events alternate. The attested cases of verbs that describe externally

caused events, but do not alternate, such as (5.6b), are considered exceptions due to some

idiosyncratic semantic properties of events described by the verbs (Section 5.2). The

quantitative notion of scale of spontaneous occurrence allows expressing the differences

between the verbs that describe externally caused events. Each point on the scale

represents a different probability for an agent to occur in an event described but a verb.

Opposite to the spontaneous extreme of the scale there is the non-spontaneous extreme.

It predicts cases of verbs describing events that are very unlikely to occur spontaneously.

An intransitive use of these verbs would be unlikely, although possible. The case in (5.6b)

could be explained in these terms with no need to treat it as an exception.

5.3. Experiments

Our approach to lexical causatives is based on statistical analysis and modelling of large

data sets. Assuming that the use of verbs is related to their semantic and grammatical

properties, we observe the distribution of the causative and anticausative realisations of a

179


large number of verbs extracted from a corpus of around 1’500’000 syntactically analysed

sentences identifying the properties of verbs which generated this distribution.

We first show that the distribution of the two realisations of the alternating verbs in

a corpus is correlated with the distribution of morphological marking across languages.

We measure the correlation for a sample of 29 verbs for which typological descriptions

are available (Haspelmath 1993). Regarding this correlation as a piece of evidence that

the two distributions are generated by the same underlying property of the alternating

verbs, we define this property as the degree of involvement of an external causer in an

event described by a verb. Following Haspelmath (1993), we call this property the degree

of spontaneity of an event. The more spontaneous an event, the less is an external causer

involved in the event. We see the degree of spontaneity as a general scalar component of

the meaning of lexical causatives whose value has an impact on the observable behaviour

of all the verbs which participate in the alternation in any language.

Showing that the corpus-based measure of spontaneity is correlated with the typological

measure allows us to extend the account to a larger sample of verbs. Since the corpus-

based value is assigned to the verbs entirely automatically, it can be quickly calculated for

practically any given set of verbs, replacing the typology-based value for which the data

are harder to collect. We calculate the corpus-based value of the spontaneity of events

described by 354 verbs cited as participating in the causative alternation in English

(Levin 1993). We show, by means of a statistical test, that the smaller set of verbs

(the 29 verbs for which we measured the correlation) is a proper sample of the bigger

set (the 354 verbs from Levin (1993)). This implies that the correlation established for

the smaller set applies to the bigger set as well.

To study how exactly the spontaneity value influences the cross-linguistic variation, we

analyse the distribution of causative and anticausative realisations in German transla-

tions of English lexical causatives. We extract the data from the corpus of German

translations of the 1’500’000 English sentences which were used in the monolingual part

of the study. The sentences on the German side are, like English sentences, syntactically

analysed. All the sentences are word-aligned so that German translations of individual

English words are known. By a statistical analysis of parallel instances of verbs, we

180

5.3. Experiments

identify certain trends in the cross-linguistic variation which are due to the spontaneity

value.

Based on these findings, we design a probabilistic model which exploits the information

about the cross-linguistic variation at the level of token to assess the spontaneity value

of lexical causatives abstracting away from the potential language-specific biases.

5.3.1. Experiment 1: Corpus-based validation of the scale of

spontaneous occurrence

Haspelmath (1993) does not discuss a potential relation between the likelihood of spon-

taneous occurrence of an event and the frequency of different uses of the verb which

describes it in a single language. Nevertheless, it is logical to suppose that such a rela-

tion should exist, since the indicator of the likelihood, the morphological marking on the

verbs, is considered to be a consequence of the way the verbs are used in general. The

placement of an event described by a verb on the scale can be expected to correspond

to the probability for the verb to be used transitively or intransitively in any single lan-

guage. On the other hand, the ratio of the frequencies of intransitive to transitive uses of

verbs in a single language can be influenced by other factors as well, which can result in

cross-linguistic variation. The relation between the scale of spontaneous occurrence and

the patterns of use of verbs in different languages thus needs to be examined. Note that

the causative alternation is realised in different ways across languages: some languages

mark the causative use of a verb, some mark the anticausative use, some mark both and

some none of them (see Tables 5.1 and 5.3). Morphological markers themselves can be

special causative morphemes, but they can often be morphemes that have other func-

tions as well, such as the reflexive anticausative marker in most of European languages.

These factors might have influence on the ratio of intransitive to transitive uses in a

given language.

To validate empirically the hypothesis that the alternating verbs can be ordered on the

scale of spontaneous occurrence of the events that they describe, we test it on corpus

data. More precisely, we test the hypothesis that the distribution of morphological

181


marking on the verbs across languages and the distribution of their transitive to intran-

sitive uses in a corpus are correlated. We can expect this correlation on the basis of the

well established correspondence between markedness and frequency. In general, marked

forms are expected to be less frequent than unmarked forms. Therefore, we expect the

verbs that tend to have anticausative marking across languages to be used more often as

causative (transitive), and verbs that tend to have causative marking to be used more

often as anticausative (intransitive). To make the discussion easier to follow, we opt for

a positive, and not a negative correlation. We thus calculate the C/A and not the A/C

ratio as Haspelmath (1993).

We calculate the ratio between the frequencies of causative (active transitive) and an-

ticausative (intransitive) uses of verbs in a corpus of English for the verbs for which

Haspelmath’s study provides the typological A/C ratio, as shown in (5.10).

C/A(verb) =frequency of causative uses

frequency of anticausative uses(5.10)

We then measure the strength of the correlation between the ranks obtained by the two

measures.


Several explanations are needed regarding the matching between the verbs in Haspel-

math’s and our study and the criteria that were used to exclude some verbs. Most of

the verbs analysed by Haspelmath are also listed as participating in the causative al-

ternation by Levin (1993) (e.g. freeze, dry, melt, open, break, close). Some verbs are

not listed by Levin (e.g. boil, gather). We include them in the calculation neverthe-

less because they clearly alternate. Four entries in Haspelmath’s list are not English

alternating verbs, but complement pairs: learn/teach, rise/raise, go out/put out, and

get lost/lose. We treat the former two pairs as single verb entries adding up counts of

occurrences of both members of the pair. We do not calculate the ratio for the latter two

because automatic extraction of their instances from the corpus could not be done using

the methods already developed to extract the other verb instances. We exclude the verb

182

5.3. Experiments

destroy because it does not alternate in English and no complement verb is proposed by

Haspelmath. Finally, the pair kill/die is excluded because its typology-based ranking is

not available. This leaves us with 27 verbs for which we calculate the corpus-based C/A

ratio.

Transitive, intransitive, and passive instances of the verbs were extracted from the En-

glish side of the parallel corpus Europarl (Koehn 2005), version 3, which contains around

1’500’000 sentences for each language (the same corpus which was used for the study on

light verb construction presented in Chapter 4). Syntactic relations needed for deter-

mining whether a verb is realised as transitive (with a direct object) or as intransitive

(without object) are identified on the basis of automatic parsing with the MaltParser, a

data-driven system for building parsers for different languages (Nivre et al. 2007).

Instance representation. Each instance is represented with the following elements:

the verb, the head of its subject, and the head of its object (if there is one). An English

causative use of a verb is identified as an alternating verb realised in an active transitive

clause. The anticausative use is identified as an intransitive use of an alternating verb.

Passive is identified as the verb used in the form of passive participle and headed by

the corresponding passive auxiliary verb. Identification of the form of the clause which

contains a lexical causative is performed automatically, using the algorithm shown in

Algorithm 2.

Regarding all the transitive uses of the alternating verbs as causatives, and intransitive

uses as anticausatives is a simplification, because this is not always true. It can happen

that a verb alternates in one sense, but not in another. For instance, the phrase in (5.12)

is not the causative alternation of (5.11), but only of (5.13).

(5.11) Mary was running in the park this morning.

(5.12) Mary was running the program again.

(5.13) The program was running again.

By a brief manual inspection of the lexical entries of the verbs in the Proposition Bank

(Palmer et al. 2005a), we assessed that this phenomenon is not very frequent and that

it should not have an important influence on the results. In our sample, only the verb

183


freeze proved to be affected by this phenomenon. This verb was discarded as an outlier

while calculating the correlation between corpus based and typology based rankings of

the verbs, but this was the only such example in the sample of verbs.

Algorithm 2: Identifying transitive, intransitive, and passive uses of lexical causatives.Input : 1. A corpus S consisting of sentences s parsed with a dependency parser

2. A list of lexical causatives V

Output : The number of transitive, intransitive, and passive instances of each verb v ∈ V in

the corpus S

for i = 1 to i = S do

for j = 1 to j = V do

if vj in si then

if there is SUBJ which depends on vj then

if there is OBJ which depends on vj then

return transitive;

else

if vj is passive then

return passive;

else

return intransitive;

end

end

end

end

end

end

As it can be seen in Algorithm 2, only the instances with all the arguments realised in

the same clause were taken into account. This is obtained by the constraint that the

extracted verb has to have a subject. We exclude the realisations of verbs where either

the subject or the object are moved or elided in order to control for potential influence

that the specific syntactic structures can have on the interpretation of the meaning of

verbs. Single clause realisations can be considered the typical and the most simple

case.

184

5.3. Experiments

Figure 5.1.: The correlation between the rankings of verbs on the scale of spontaneousoccurrence

Although they are basically transitive realisations, the passive instances are extracted

separately because the difference between active and passive transitive uses is crucial

with respect to the causative alternation, as discussed in detail by Alexiadou et al.

(2006). Expressing the external causer (by means of a prepositional complement) is

optional in passive constructions, while in active transitive instance, the external causer

is obligatorily expressed as the subject.


To asses the strength of correlation between the corpus-based C/A ratio and the A/C

ratio based on the typology of morphological marking on the verbs, we rank the verbs ac-

cording to the corpus use ratio and then perform a correlation test between the rankings

of the same verbs based on the two measures. We obtain the Spearman rank correlation

185


score rs = 0.67, p < 0.01, with one outlier7 removed. The score suggests good correla-

tion between the two sources of data. Figure 5.1 shows the scattergram representing the

correlation.

The coefficient of the correlation is strong enough to be taken as an empirical confir-

mation of Haspelmath’s hypothesis. Given that the two distributions are significantly

correlated, it is reasonable to assume that the same factor which underlies the typolog-

ical distribution of morphological marking on verbs underlies the distribution of their

transitive and intransitive realisations in a monolingual corpus too. Since the correlation

is established based on the intuition that the underlying cause of the observed distribu-

tions is the meaning of verbs, we can conclude that the lexical property on which the

two distributions depend is the probability of occurrence of an external causer in an

event described by a verb.

5.3.2. Experiment 2: Scaling up

The fact that automatically obtained corpus-based ranking of verbs corresponds to the

scale of spontaneous occurrence is a useful finding, not only because it confirms Haspel-

math’s theoretical hypothesis, but also because it means that the spontaneity feature

can be calculated automatically from corpus data. In this way, it is possible to extend

the account beyond the small group of example verbs that are discussed in the literature

and cover many more cases. To test whether the correlation that we find for the small

sample of verbs discussed in Section 5.3.1 applies to a larger set, we compare the distri-

bution of the corpus-based measure of spontaneity over this sample and the distribution

of the same value for the 354 verbs listed by Levin (1993) (see Section 2.2.1 in Chapter

2 for more details).

7Verb freeze is frequently used in our corpus in its non-literal sense (e.g. freeze pensions, freeze assets),while the sense that was taken into account by Haspelmath (1993) is most likely the literal meaningof the verb (as in The lake freezed.). This is why the verb’s corpus-based ranking was very differentfrom its typology based ranking.

186

5.3. Experiments


The list of English lexical causatives is extracted from the (Levin 1993) verb index. Since

the index referred with the same number to the verbs that do not enter the alternation

(the book sections 1.1.2.1, 1.1.2.2, and 1.1.2.3), the verbs that do not alternate were

removed from the list manually.

All the instances where these verbs occur as transitive, intransitive, and passive were

extracted from the automatically parsed English side of the Europarl corpus. We extract

the same counts which were extracted for the small sample discussed in Section 5.3.1.

We reduce the variance in the corpus-based measure preserving the information about

the ordering of verbs by transforming the frequency ratio into a more limited measure of

spontaneity. We calculate the value of spontaneity (Sp in 5.14) for each verb v included

in the study as the logarithm of the ratio between the rates of anticausative and causative

uses of the verb in the corpus, as shown in (5.14).

Sp = ln

(rate(v, caus)

rate(v, acaus)

)(5.14)

The rates of uses of the three extracted constructions form ∈ {anticaus, caus, pass} for

each verb are calculated as in (5.15).

rate(form) =F (form, v)∑

form F (form, v)(5.15)

The verbs that tend to be used as anticausative will have negative values for the variable

Sp, the verbs that tend to be used as causative will be represented with positive values,

and those that are used equally as anticausative and causative will have values close to

zero. The distribution of the Sp-value over the 354 verbs is shown in Figure 5.4.

In the cases of verbs that were not observed in one of the three forms, we calculated the

rate values as the rate of uses of the form in the instances of all verbs with frequency one

divided by the total frequency of the verb in question. For example, the verb attenuate

occurred three times in the corpus, once as causative, and two times as passive. The

187


Figure 5.2.: Density distribution of the Sp value in the two samples of verbs

rate of anticausative uses for this verb is 0.31/3 = 0.10. The number 0.31 that is used

instead of the observed count 0 represents the rate of all verbs with frequency one that

occurred as intransitive. After normalising, the rate of causative uses of this verb is 0.30.

The rate of passive uses is 0.61, and the rate of anticausative uses is 0.09. In this way

we obtain small non-zero values proportional to the two observed frequencies.


We compare the distribution of the Sp-values over the small and the large set of verbs

in several ways. Figure 5.2 shows the density distribution of the spontaneity value over

the two samples of verbs. First, visual assessment of the shapes of the two distributions

suggests that they are very similar. They both have a single mode (a single most probable

value). Both modes are situated in the same region (around the value 0). The difference

in the probability of the most probable values which can be observed in the figure (0.6

188

5.3. Experiments

for the large sample as opposed to 0.3 for the small sample) does not necessarily reflect

the real difference in the two distributions. It can be explained by the fact that the

large sample contains a number of unobserved verbs for which are assigned the same Sp-

value, estimated on the basis of the values of low frequency verbs, as discussed earlier.

In reality, the verbs would not have exactly the same value, so that the density would be

more equally distributed around zero, which is exactly the case in the small sample.

Another indication that the two samples are the same is the value of two-vector t-test

t = −0.0907, p = 0.9283, which indicates a very small difference in the means of the two

distribution and a high probability that it is observed by chance. The t-test works with

the assumption that the distributions which are compared belong to the family of normal

distributions. It does not apply to other kinds of distributions. To make sure that the

distributions of our data can be compared with the t-test, we perform the Shapiro-Wilk

test which shows how much a distribution deviates from a normal distribution. This

test was not significant (W = 0.9355, p = 0.07), which means that the distribution of

our data can be considered normal.

We conclude that the verbs for which the corpus-based ranking is shown to correspond

to the typology-based ranking represent an unbiased sample of the larger population of

verbs that participate in the causative alternation. This implies that the corpus-based

method for calculating the spontaneity value presented in this section can be applied to

all the verbs that participate in the alternation.

The limitation of this method, however, is that it is based on the observations in a mono-

lingual corpus. Given the well-documented cross-linguistic variation in the behaviour of

the alternating verbs, discussed in Section 5.2.2 and 5.2.3, and summarised in Table 5.1,

a monolingual measure is likely to be influenced by language-specific biases in the data.

In the following section, we take a closer look at the relationship between the patterns

of cross-linguistic variation in the instances of lexical causatives and their spontaneity

value.

189


5.3.3. Experiment 3: Spontaneity and cross-linguistic variation

In analysing cross-linguistic variation in the realisations of lexical causatives, we try to

determine whether a verb can be expected to have consistent or inconsistent realisations

across languages depending on the degree to which an external causer is involved in the

event described by the verb.

We approach this task by analysing German translations of English lexical causatives

as they are found in a parallel corpus. Studying instances of translations of lexical

causatives in a parallel corpus allows us to control for any pragmatical and contextual

factors that may be involved in a particular realisation of a lexical causative. Since

translation is supposed to express the same meaning in the same context, we can assume

that the same factors that influence a particular realisation of a verb in a clause in one

language influence the realisation of its translation in the corresponding clause in another

language. Any potential differences in the form of the two parallel clauses should be

explained by the lexical properties of the verbs or by structural differences between the

languages.

We perform a statistical analysis of a sample of parallel instances of lexical causatives

in English and German, which we divide into three subsamples: expressions of sponta-

neous events, expressions of non-spontaneous events and expressions of the events that

are neutral with respect to spontaneity. Given that spontaneity of an event, as a univer-

sal property, correlates with causative and anticausative use monolingually, and given

that translations are meaning-preserving, we expect to find an interaction between the

level of spontaneity of the event described by the verb and its cross-linguistic syntactic

realisation. Assuming that cross-linguistic variation is an extension of within-language

variation, as discussed in the end of Section 4.3 in Chapter 4, we expect syntactic reali-

sations consistent with the lexical semantics of the verb to be carried across languages in

a parallel fashion, while those that are inconsistent are expected to show a tendency to-

wards the consistent realisation. For example, we expect intransitive realisations to stay

intransitive, and transitives to be often transformed into intransitives when verbs de-

scribe spontaneous events. Since the probability of both realisations is similar in neutral

instances, we expect to find fewer transformations than in the other two groups.

190

5.3. Experiments


The data collected for this analysis comes from large and complex resources. To know

which English form is aligned with which German form, we first need to extract the

English lexical causative from the English side of the parallel corpus. We then determine

its form based on the automatic syntactic analysis of the sentence. Once we know which

sentence in the English side of the parallel corpus contains a lexical causative, we find the

German sentence which is aligned with the English sentence based on automatic sentence

alignment in the parallel corpus. Once the aligned German sentence is identified, we

look for the German verb which is aligned with the English verb. To do this, we first

search the automatic word-alignments to find the German word which is aligned with

the English verb. If we find such a word, we then look into the syntactic analysis of the

German sentence to determine whether this word is a verb. If it is a verb, we then search

the German syntactic parse to find the constituents of the clause where the verb is found.

Once we know the constituents, we can determine whether the German translation of

the English lexical causative is transitive, intransitive or passive (using the same criteria

as for extracting English instances in Section 5.3.2). The methods used to collect the

data are described in more detail in the following subsection.

The verbs included in this study are the 354 English verbs listed as alternating by Levin

(1993), for which we have calculated the Sp-value applying the procedure described in

Section 5.3.2. We extract the parallel instances of these verbs from a parallel English-

German Europarl (Koehn 2005) corpus, version 3 (the same corpus which is used in the

study in Chapter 4). The corpus consists of German translations of around 1’500’000

English sentences which are used in the previous two experiments (see Section 5.3.1).

Note that by German translation we mean German translation equivalents, since the

direction of translation is not known for most of the corpus.

To extract the information about the syntactic form of the instances, needed for our

research, the German side of the corpus is syntactically parsed using the same parser as

for the English side, the MaltParser (Nivre et al. 2007). The corpus was word-aligned

using the system GIZA++ (Och and Ney 2003) (the same tool which is used in the

study in Chapter 4.) Both for the syntactic parses and word alignments are provided by

Bouma et al. (2010) who used these tools to process the corpus for their own research.

191


We extract the data for our research reusing the processed corpus (with some adaptation

and conversion of the format of the data).

In extracting the data for our analysis, we search the processed parallel corpus looking

for four pieces of information: English syntactic parse, alignment between English and

German sentences, alignment between English and German words, and German syntac-

tic parse. All four pieces of information are not available for all the sentences in the

processed resource. Some English sentences are syntactically analysed, but the corpus

does not contain their translations to German. Likewise, there are German sentences for

which English translations are not available. Finally, syntactic parses are not available

for all the sentences which are aligned. Having established these mismatches, we first

search the whole resource to find the items which contain all the required information.

Once we have found the intersection between the English parses, the German parses and

the sentence alignment, we search the English side of these sentences to identify English

lexical causatives in the same way as in Section 5.3.2.

The German translation of each instance of an English lexical causative is extracted on

the basis of word alignments. Instances where at least one German element was word-

aligned with at least one element in the English instance were considered aligned. The

extraction procedure is shown in more detail in Algorithm 3.

192

5.3. Experiments

Levin (1993)verb index

Parsed EnglishEuroparl,Prolog format

Parsed GermanEuroparl,Prolog format

List of lexicalcausatives

Parsed EnglishEuroparl,CoNLL format

Parsed GermanEuroparl,CoNLL format

Extract Englishinstances

Englishcausativescounts

Englishinstances

Word alignedEnglish-GermanEuroparl, Prologformat

Corpus-based spontaneitymeasure

Extractwordalignmentof Englishinstances

Correctedwordalignment

Experimentaldata set

Figure 5.3.: Data collecting workflow. The shaded boxes represent external resources.The dashed boxes represent the scripts which are written for specific tasks inextracting data and performing the calculations. The other boxes representthe input and the output data at each stage of data processing.

193


Algorithm 3: Extracting parallel cross-linguistic realisations of lexical causatives.Input : 1. A corpus E consisting of English sentences e which

a. contain realisations of lexical causatives

b. are parsed with a dependency parser

c. annotated with the form of the realisation (transitive, intransitive, or passive)

-

2. A corpus G consisting of German sentences g which are

a. sentence- and word-aligned with English sentences in E

b. parsed with a dependency parser.

Output : Parallel instances consisting of:

a. the form of the English realisation (transitive, intransitive, or passive)

b. the form of the aligned German realisation (transitive, intransitive, or passive)

for i = 1 to i = E do

if align(verbj) is a German verb then

g-verb = align(verbj);

do Algorithm-2(g-verb)

else

if there is align(OBJj) then

g-verb = the verb on which align(OBJj) depends;


else

if there is align(SUBJj) then

g-verb = the verb on which align(SUBJj) depends;


else

return no align;

end

end

end

end

We define the alignment in this way to address the issue of missing alignments. As

discussed in more detail in Section 4.4 in Chapter 4, the evaluation of the performance

of the word-alignment system GIZA++ on the Europarl data for English and German

(Pado 2007) showed a recall rate of only 52.9%, while the precision is very high (98.6%).

194

5.3. Experiments

This evaluation applies to the intersection of both directions of word alignment, which is

the setting used in the processing of our data. The low recall rate means that around half

of the word alignments are not identified by the system. Extracting only the instances

where there is a word alignment between an English and a German verb would hence

considerably reduce our data set. Instead, we rely on the extracted syntactic relations

and on the intuition that the verbs are aligned if any of the constituents which depend

on them are aligned. With this definition of instance alignment, we also take advantage

of our own finding that nouns (the heads of the depending constituents are nouns) are

generally better aligned than verbs, as discussed in Chapter 4.

Sentence ID Verb Form Verb instance Subject Object

96-11-14.867 intensify CAUS 7 5 8

96-11-14.859 beschleunigen CAUS 6 2 5

Table 5.4.: An example of an extracted instance of an English alternating verb and itstranslation to German. The numbers under the elements of the realisationsof the verbs indicate their position in the sentence. For example, the objectof the English verb is the eight word in the sentence 96-11-14.867, and theobject of the German verb is the fifth word in the sentence 96-11-14.859.

A pair of extracted aligned instances is illustrated in Table 5.4. The first column is the

sentence identifier, the second column is the verb found in the instance, the third column

is the form of the verb in the instance, and the following three columns are the positions

of the verb, the head of its subject and the head of its object in the sentence.

One more processing step was needed to identify sentence constituents which are word-

aligned because the word alignments and the syntactic analysis did not refer to the same

positions. This is caused by the fact that sentence alignment was often not one-to-one.

In the cases where more than one English sentence were aligned with a single German

sentence, or the other way around, the positions of words were determined with respect

to the alignment chunk, and not with respect to the individual sentences. For example,

if two English sentences were aligned with a German sentence, with eight words in the

first sentences and seven in the second. The position of the first word in the second

195


Sp En De1.20 pass intrans1.97 trans trans0.71 trans trans

-0.05 pass pass0.71 trans trans

-0.09 trans pass-0.14 trans intrans-3.91 intrans intrans0.39 pass intrans

-1.76 intrans trans

Table 5.5.: Examples of parallel instances of lexical causatives.

sentence is indicated as 9. In the syntactic parse, on the other hand, these two sentences

are not grouped together, so the position of the same word is indicated as 1. We restored

the original sentence-based word enumeration in the word alignments before extracting

the alignment of the constituents.

Applying the described methods allows us to extract only translations with limited

cross-linguistic variation. Only the instances of English verbs that are translated with

a corresponding finite verb form in German are extracted, excluding the cases where

English verbs are translated into German with a corresponding non-finite form such as

infinitive, nominalization, or participle in German.

The extracted parallel instance were then combined with the information about the Sp-

value for each verb to form the final data source for our study, as illustrated in Table

5.5. Each line represents one instance of an alternating English verb found in the corpus

and its translation to German. The full data set contains 13033 such items. The first

column contains the spontaneity value of the verb found in the instance. The second

column represents the form in which the English verb is realised in the instance. The

third column represents the form of the German translation of the English verb. Figure

5.3 shows the main steps in the data collecting work flow.

196

5.3. Experiments

Validation of the collected data. Since all the data used in our study are collected

automatically from an automatically parsed and word-aligned corpus, they necessarily

include processing-related errors. The best reported labelled attachment score of the

MaltParser system for English is 88.11 (CoNLL Shared Task 2007) and for German 85.82

(CoNLL-X Shared Task 2006).We perform a manual evaluation of a sample of randomly

selected instances to asses to what degree they correspond to the actual analyses.

One hundred parallel instances were randomly selected from the total of 13033 extracted

instances. The following categories were evaluated:

• The form of the clause in the English instance

• The form of the clause of the German translation

The extraction script assigned a wrong form to 8/100 English instances (error rate 8%).

In 7 cases out of 8 errors, the wrong form was assigned due to parsing errors. One error

was due to the fact that the parser’s output does not include information about traces.

For example, in a sentence such as That is something we must change, the anticausative

form is assigned to the instance of the verb change instead of the causative form. In four

out of the seven parsing errors the actual forms found in the instances were not verbs

but adjectives (open, close, clear, worry).

The evaluation of the translation extraction was performed only for the cases where the

English instance actually contained a verbal form (96 instances). A wrong form was

assigned to the German translation in 13/96 cases (error rate 13.5%). In 7 of the 13

wrong assignments, a wrong form was assigned to the translation due to parsing errors

in German. The errors in 3 cases were due to the fact that German passive forms headed

by the verb sein, as in Das Team war gespalten for English The team was split, were not

recognised as passive, but they were identified as anticausative instead. The ambiguity

between such forms and anticausative past tense formed with the sein auxilliary verb

cannot be resolved in our current extraction method. In the last 3 cases, the error

was due to the fact that the corresponding German form was not a clause. In these

cases, the English verb is aligned to a word with a different category (an adverb and a

nominalization) or entirely left out in the German sentence (a verb such as sit in We sit

here and produce...). The form that is assigned to the translation in these cases is the

197


Figure 5.4.: Density distribution of the Sp value over instances of 354 verbs

form of the verb on which the aligned words depend. Our extraction method cannot deal

with these cases at the moment, although such transformations would be interesting to

capture.

Sampling of instances. Three groups of instances are defined according to the density

distribution of the Sp value. As it can be seen in Figure 5.4, roughly symmetric points of

low density are around values -1 and 1. We regard the instances containing the verbs with

the value of Sp inferior to -1 as the low value group. These are expressions of spontaneous

events in the terms of the scale of spontaneous occurrence (see Section 5.2.3), or of

internally caused events in the sense of the theories presented in Section 5.2.1 and 5.2.2.

Instances containing a verb with the Sp value superior to 1 are considered belonging to

the high value group, representing expressions of non-spontaneous, or externally-caused

events. The instances in between the two values are considered medium value instances,

198

5.3. Experiments

representing what Alexiadou (2010) refers to as cause unspecified events (see Section

5.2.2).

This division gives symmetric sub-samples of comparable size: similar number of exam-

ples for the two extreme values (3’107 instances with high Sp values, 2’822 instances with

low Sp values) and roughly double this number for non-extreme values (7’104 instances

with medium Sp values).


Table 5.6 shows the frequencies of the realisations of lexical causatives in parallel En-

glish and German instances for the whole sample of instances, as well as for the three

sub-samples. The three most frequent combinations of forms in each group of parallel

instances are highlighted to show the changes in the distribution of combinations of

forms in the two languages across groups.

The overview of the frequencies suggests that lexical properties of verbs influence their

cross-linguistic realisations.

The table that shows occurrences over the whole sample indicates that, both in English

and in German, intransitives are more frequent than transitives, which are, in turn, more

frequent than passives (marginal distributions). The non-parallel translations cover 32%

of the cases.

When we partition the occurrences by the spontaneity of the event, the distribution

changes, despite the fact that these are distributions in translations, and therefore sub-

ject to very strong pressure in favour of parallel constructions.

In the group of instances containing verbs that describe events around the middle of the

scale of spontaneous occurrences, the parallel combinations are the most frequent, as

in the distribution of the whole set, with an even more markedly uniform distribution

(29% of non-parallel translations). This means that the verbs which describe events

which are neither spontaneous or non-spontaneous tend to be used in the same form

across languages. The probabilities of the two realisations are similar in these verbs,

which means that they can be expected to occur with similar frequency across languages.

199


WholeGerman

Intransitive Transitive Passive Totalsample N % N % N % N %

Engl

ish Intransitive 3504 27 1001 8 314 2 4819 37

Transitive 1186 9 2792 21 369 3 4347 33Passive 781 6 517 4 2569 20 3867 30Total 5471 42 4310 33 3252 25 13033 100

SpontaneousGerman

Intransitive Transitive Passive Totalevents N % N % N % N %

Engl



Non- Germanspontaneous Intransitive Transitive Passive Total

events N % N % N % N %

Engl



NeutralGerman

Intransitive Transitive Passive Totalevents N % N % N % N %

Engl



Table 5.6.: Contingency tables for the English and German forms in different samples ofparallel instances.

Since both realisations are frequent in these verbs, they can be expected to alternate in

the majority of languages.

The distribution of the forms is different in the groups of instances containing verbs

that describe events on the extremes of the scale of spontaneous occurrence. The parallel

200

5.3. Experiments

Figure 5.5.: Joint distribution of verb instances in the parallel corpus. The size of theboxes in the table represents the proportion of parallel instances in eachsub-sample.

realisations are frequent only for the forms that are consistent with the lexical properties

(intransitive for spontaneous events and transitive for non-spontaneous events). An

atypical instance of a verb in one language (e.g. transitive instance of a verb that

describes a spontaneous event) is not preserved across languages. These realisations

tend to be transformed into the typical form in another language. For example, German

transitives are much less frequent in the spontaneous events group than in the non-

spontaneous events group, while English intransitive non-spontaneous verbs are only

5% compared to 82% of the spontaneous group. The atypical realisations of these verbs

are thus rare across languages, which means that they might be entirely absent in some

languages. In the languages in which these realisations are found, the verbs alternate,

while in the languages where these realisations are not found the verbs do not alternate.

This means that the verbs describing events on the extremes of the scale of spontaneous

occurrence can be expected to alternate in a smaller range of languages.

We conclude that the analysis of the realisations of lexical causatives in a parallel corpus

provides evidence that the probability of occurrence of an external cause in the event

described by a verb (the spontaneity of the event) is a grammatically relevant lexical

property. The cross-linguistic variation in the availability of the alternation is influenced

by this property. Verbs that describe events on the extremes of the scale of spontaneous

occurrence are more likely to have different realisations across languages than those that

describe events in the middle of the scale.

201


5.3.4. Experiment 4: Learning spontaneity with a probabilistic

model

In Section 5.3.1 and Section 5.3.2, we have shown that the spontaneity value of verbs

can be estimated from the information about the distribution of the causative and an-

ticausative instances of verbs in a corpus. These estimations, however, are based on

the data from only one language. Given that realisations of the alternation in different

languages is influenced by unknown factors, resulting in the observed cross-linguistic

variation, the estimation based on the data from a single language can be expected

to be influenced by language-specific factors. As we could see in Section 5.3.1 the es-

timation of the spontaneity value based on English corpus data is correlated with an

estimation based on the data from many different languages. The correlation, however,

is not perfect and the reason for this could be the deviation of the realisations in English

from the general tendencies. For example, it has been established that English prefers

causative realisations of some verbs compared to other languages (Bowerman and Croft

2008; Wolff et al. 2009a; Wolff and Ventura 2009b). As a result, estimations based on

English data can give lower spontaneity values compared to the universal value.

Another indication of potential language-specific factors which influence the way lexical

causatives are realised in a language can be found in the results of the experiment

presented in Section 5.3.3. While this experiment shows that the spontaneity value

influences the cross-linguistic realisations, we can also see that there is a number of

realisations which are divergent across languages. As discussed in Section 5.3.3, the

realisations which are not consistent with the spontaneity value in one language tend

to be transformed into realisations consistent with the spontaneity value in the other

language. However, the factors which give rise to the realisations inconsistent with the

spontaneity are not known.

To address the issue of potential influence of language-specific factors on corpus-based

estimation of the spontaneity value of events described by alternating verbs, we extend

the corpus based approach to the cross-lingsuitc domain. We collect the information

about the realisations of the alternating verbs in a parallel corpus, as described in Section

5.3.3. The extended data set is expected to provide a better estimation of the universal

spontaneity value than the monolingual set, neutralising the language-specific influences.

202

5.3. Experiments

Naturally, including more languages would be expected to give even better estimates.

In this study, however, we consider only two languages as a first step towards a richer

cross-linguistic setting.

A simple approach to integrating cross-linguistic corpus data would be to calculate the

ratio of causative to anticausative uses by adding up the counts from different languages.

For instance, if a verb is found as transitive four times in one language and the trans-

lations of these four instances were two times transitive and two times intransitive, the

count of transitive instance of these verb would be six. The two intransitive translations

would be added to the counts of intransitive instances. In this approach, however, the

information about which instances were observed in which language is lost. This knowl-

edge, however, can be very important for isolating language-specific factors and using

this information to predict cross-linguistic transformations of phrases containing verbs

with causative meaning. Another disadvantage of such an approach, which applies to

all the estimations performed in this study so far, is that they do not provide a straight-

forward way of grouping the verbs into classes, which is one of the major concerns in

the representation of lexical knowledge (see Section 5.2 in this chapter and also Section

2.2.1 in Chapter 2).

To take into account both potential language specific factors and potential grouping of

verbs we design a probabilistic model which estimates the spontaneity value on the basis

of cross-linguistic data and generates a probability distribution over a given number of

spontaneity classes for each verb in a given set of verbs.

The number of classes. Two main proposals concerning the classification of alternat-

ing verbs have been put forward in the linguistic literature. As discussed in Section 5.2,

Levin and Rappaport Hovav (1994) use the distinction between externally and internally

caused events to explain a set of observations concerning the alternating verbs. Alexi-

adou (2010), however, points out that a range of cross-linguistic phenomena are better

explained by introducing a third semantic class, the cause-unspecified verbs. The dis-

tinctions argued for in the linguistic literature can be roughly related to the spontaneity

feature in our account, so that externally caused events correspond to non-spontaneous,

internally caused to spontaneous and cause unspecified to medium-spontaneity events.

203


The model

As it can be seen in its graphical representation in Figure 5.6, the model consists of four

variables.

V

Sp

En Ge

Figure 5.6.: Bayesian net model for learning spontaneity.

The first variable is the set of considered verbs V . This can be any given set of verbs.

The second variable is the spontaneity class of the verb, for which we use the symbol

Sp. The values of this variable depend on the assumed classification.

The third (En) and the fourth (Ge) variables are the surface realisations of the verbs in

parallel instances. These variables take three values: causative for active transitive use,

anticausative for intransitive use, and passive for passive use.

We represent the relations between the variables by constructing a Bayesian network

(for more details on Bayesian networks, see Section 3.4.3 in Chapter 3), shown in Figure

5.6. The variable that represents the spontaneity class of verbs (Sp) is treated as an

unobserved variable. The values for the other three variables are observed in the data

source. Note that the input to the model, unlike the information extracted for the anal-

ysis in Section 5.3.3, does not contain the information about the spontaneity (compare

Table 5.7 with Table 5.5).

204

5.3. Experiments

The dependence between En and Ge represents the fact that the two instances of a

verb are translations of each other, but does not represent the direction of translation.

The form of the instance in one language depends on the form of the parallel instance

because they express the same meaning in the same context, regardless of the direction

of translation.

Assuming that the variables are related as in Figure 5.6, En and Ge are conditionally

independent of V given Sp, so we can calculate the probability of the model as in

(5.16).

P (v, sp, en, ge) = P (v) · P (sp|v) · P (en|sp) · P (ge|sp, en) (5.16)

Since the value of spontaneity is not observed, the parameters of the model that involve

this value need to be estimated so that the probability of the whole model is maximised.

We estimate the Sp-value for each instance of a verb by querying the model, as shown

in (5.17).

P (sp|v, en, ge) =P (v, sp, en, ge)∑sp P (v, sp, en, ge)

(5.17)

Applying the independence relations defined in the Bayesian net (Figure 5.6), the most

probable spontaneity class in each instance is calculated as shown in (5.18).

sp = arg maxsp

P (v) · P (en|sp) · p(ge|sp, en)∑sp P (v) · P (en|sp) · p(ge|sp, en)

(5.18)

Having estimated the Sp-value for each verb instance, we assign to each verb the average

spontaneity value across instances, as shown in (5.19).

sp class(verb) =

∑en

∑ge p(sp|v, en, ge)F (v)

(5.19)

where F (v) is the number of occurrences of the verb in the training data.

205


All the variables in the model are defined so that the parameters can be estimated

on the basis of frequencies of instances of verbs automatically extracted from parsed

corpora. The corpus used as input does not need to be annotated with classes, since the

parameters are estimated treating the class variable as unobserved.

The model described in this section includes only two languages because we apply it to

the two languages that we choose as a minimal pair (English and German), but it can

be easily extended to include any number of languages.

Experimental evaluation

The accuracy of the predictions of the model is evaluated in an experiment. We imple-

ment a classifier based on the model, which we train and test using the data extracted

from the syntactically parsed and word-aligned parallel corpus of English and German,

as described in Section 5.3.3. To address the discussion on the number of classes of

alternating verbs (see Section 5.2.2), we test two versions of the model. In one version,

the model performs a two-way classification which corresponds to the binary distinction

between externally and internally caused events. In the other version, the model per-

forms a three-way classification which corresponds to the distinction between internally

caused, externally caused, and cause-unspecified events.

The verbs for which we calculate the spontaneity class in the experimental evaluation

of the model are the 354 verbs that participate in the causative alternation in English,

as listed in Levin (1993).

We estimate the parameters of the model by implementing an expectation-maximisation

algorithm, which we run for 100 iterations. (Using the algorithm to estimate the prob-

abilities in a Bayesian model are explained in Section 3.4.2 in Chapter 3.) We initialise

the algorithm according to the available knowledge about the parameters. The proba-

bility P (v) is set to the prior probability of each verb estimated as the relative frequency

of the verb in the corpus. The probability P (sp|v) is set so that causative events are

slightly more probable than anticausative events in the two-way classification, and so

that that cause-unspecified events are slightly more probable than the other two kinds

206

5.3. Experiments

V En Demove pass intransalter trans transimprove trans transincrease pass passimprove trans transbreak trans passchange trans intransgrow intrans intransclose pass intranssplit intrans trans

Table 5.7.: Examples of the cross-linguistic input data

of events in the three-way classification. The values of P (en|sp) and P (ge|sp, en) are

initialised randomly.

For the set of verbs for which the typological information is available, we compare

the classification of verbs learned by the model both with the typology-based ranking

and with the rankings based on the monolingual corpus-based Sp-value, automatically

calculated in the second experiment (see Section 5.3.2 for more details). Since the set of

verbs for which it is possible to perform a direct evaluation against typological data is

relatively small (the data for 26 verbs are available), we measure the agreement between

the classification learned by the model and the rankings based on the monolingual corpus-

based Sp-value for the set of verbs which are found in the parallel corpus (203 verbs).

This measure is expected to provide an indirect assessment of how distinguished the

supposed classes of verbs are.

Table 5.8 shows all the classifications performed automatically in comparison with the

classifications based on the typology rankings. Since the typology-based and the mono-

lingual corpus-based measures do not classify, but only rank the verbs, the classes based

on these two measures are obtained by dividing the ranked verbs according to arbi-

trary thresholds. The thresholds for classifying the verbs according to the monolingual

corpus-based Sp-value are determined in the same way as in the third experiment (see

Section 5.3.3). In the two-way classification, the threshold is Sp = −1. The verbs with

207


Verb Two-way classification Three-way classificationMonolingual bilingual. Monolingual bilingual

boil a a a adry a a a awake up a a a asink a a a alearn-teach a a m mmelt a a a astop a a a aturn a a m mdissolve c c c cburn c c c mfill c c c mfinish a a a abegin a a a aspread a a a aroll c c m mdevelop c c m mrise-raise c c c cimprove c c m mrock c c m mconnect c c c cchange a a a agather c c m mopen c c m mbreak c c c cclose c c c csplit c c c c

Agreement 85 % 85% 61% 69%

Table 5.8.: Agreement between corpus-based and typology-based classification of verbs.The classes are denoted in the following way: a=anticausative (interanallycaused), c=causative (externally caused) , m=cause-unspecified.

208

5.3. Experiments

the Sp-value below -1 are considered anticausative, the other verbs are causative. In the

three-way classification the causative class is split into two classes using the threshold

Sp = 1. The verbs with the Sp-value between -1 and 1 are considered cause-unspecified,

while the verbs with the Sp-value above 1 are causative.

The thresholds for classifying the verbs according to the typology-based ranking are

determined for each evaluation separately so that the agreement is maximised. For

example, threshold is set after the verb turn in the first two columns of Table 5.8.

All the verbs ranked higher than turn are considered anitcausative. The others are

causative.

In the two-way classification, the two versions of the model, with monolingual and with

bilingual input, result in identical classifications. The agreement of the models with

the typological ranking can be considered very good (85%). The optimal threshold

divides the verbs into two asymmetric classes: eight verbs in the internally caused class

and eighteen in the externally caused class. The agreement is better for the internally

caused class.

In the three way-classification, the performance of both versions of the model drops.

In this setting, the output of the two versions differs: there are two verbs which are

classified as externally caused by the monolingual version and as cause-unspecified by

the bilingual version, which results in a slightly better performance of the bilingual

version. Given the small number of evaluated verbs, however, this tendency cannot be

considered significant.

The three-way classification seems more difficult for both methods. The difficulty is

not only due to the number of classes, but also to the fact that two classes are not

well-distinguished in the data. While the class of anticausative verbs is relatively easily

distinguished (small number of errors in all classifications), the classes of causative and

cause-unspecified verbs are hard to distinguish. This finding supports the two-way clas-

sification argued for in the literature. However, the classification performed by the model

indicates that the distinction between causative and cause-unspecified verbs might still

exist. Compared to the classification based on monolingual Sp-value, more verbs are

classified as cause-unspecified, and they are more densely distributed on the typological

scale. Since the model takes into account cross-linguistic variation in the realisations of

209


Monolingual Sp classAnticausative Causative

Parallel Anticausative 64 13Sp class Causative 14 112

Table 5.9.: Confusion matrix for the monolingual corpus-based measure of spontaneityand the 2-class bi-lingual model classification for 203 verbs found in theparallel corpus.

Monolingual Sp classAnticausative Unspecified Causative

Parallel Anticausative 52 19 0Sp class Unspecified 1 32 35

Causative 6 24 34

Table 5.10.: Confusion matrix the monolingual corpus-based measure of spontaneity andthe 3-class bi-lingual model classification for 203 verbs found in the parallelcorpus.

lexical causatives, the observed difference in performance could be interpreted as a sign

that the distinction between cause-unspecified and causative verbs does emerge in the

cross-linguistic context.

The performance of the two automatic classifiers on all the alternating verbs found in

the parallel corpus is compared in Tables 5.9 and 5.10. The agreement between the two

automatic methods is 87% in the two-class setting and 58 % in the three-class setting.

Again, the class of anticausative verbs is rarely confused with the other two classes, while

causative and cause-unspecified verbs are frequently confused (the agreement between

the two classifications is at the chance level). The lack of agreement between the two

methods, however, does not necessarily mean that the two classes are not distinguishable.

It can also mean that the bi-lingual probabilistic model distinguishes between the two

classes better than the monolingual ratio-based measure. The direct comparison of the

two methods with the typology scale points in the same direction.

210


Figure 5.7.: The relationship between the syntactic realisations, the morphological formand the meaning of lexical causatives.


The experiments performed in our study of morpho-syntactic realisations of lexical

causatives relate various factors involved in the causative alternation. We have estab-

lished a statistically significant correlation between frequency distribution of syntactic

alternants in a monolingual corpus and frequency distribution of morphological marking

on the verbs which participate in the causative alternation across languages. The verbs

which tend to be used in intransitive clauses tend to bear causative marking across lan-

guages. The verbs which tend to be used in transitive clauses tend to bear anticausative

marking across languages.

This finding suggests that the underlying cause of both distributions is the meaning of

verbs, as illustrated in Figure 5.7. The fact that a verb which describes an event in

which an external causer is very unlikely could be the reason why the verb occurs in in-

transitive clauses and why it bears causative marking across languages. The intransitive

realisations could be due to the fact that the meaning of the verb is such that an external

causer does not need to be expressed in most of its uses. This further implies that only

one argument of the verb is expressed and that the opposition between the subject and

the object is not needed, which gives rise to an intransitive structure. However, there is

still a possibility for such a verb to be realised so that the external causer of the event is

expressed (typically in a transitive clause). The realisations which are default for these

verbs (intransitive) are not morphologically marked because they correspond to the gen-

eral expectation. The realisations with an explicit external causer are morphologically

211


marked because they are unexpected. For the same reasons the verbs which describe

events in which an external causer is very likely tend to occur in transitive clauses, but

when they are used as intransitive, they tend to be morphologically marked. Although

we have not excluded other possible explanations, the likelihood of an external causer

in an event described by a verb seems to be a plausible underlying cause of the observed

correlation.

5.4.1. The scale of external causation and the classes of verbs

The results of our experiments suggest that the alternating verbs are spread on a scale

of the likelihood of external causation. The distribution of the corpus-based measure

of the likelihood of external causation (the term Sp-value is used in our experiments)

over a large number of verb instances in a corpus is normal, which implies that the

most likely value is the mean value, and that both extremely low and extremely high

values are equally likely. This finding suggests that most of the alternating verbs can be

expected to describe events in which there is around 50% chance for an external causer

to occur. These are the verbs which alternate in the majority of languages. However,

the probability of an external causer can be very small for some verbs. These are the

verbs which alternate only in some languages, while they do not alternate in the majority

of languages. The same can be claimed for the verbs describing events with extremely

likely external causers. Thus, the likelihood of external causation in the meaning of

verbs explains the observed cross-linguistic variation in the availability of the causative

alternation. The verbs which do not alternate in English can be expected not to alternate

in a number of other languages too. The number of languages in which a verb can be

expected not to alternate can be predicted from the likelihood of external causation in

the event which it describes.

Although the scale of likelihood of external causation seems to be continuous based

on the fact that many different verbs are assigned distinct values, the results of our

experiments on classification suggest that some values can be grouped together. The

anticausative part of the scale seems to be distinguishable from the rest of the scale.

The verbs classified as anticausative in our experiments can be related to verbs describ-

ing internally-caused events discussed in the literature. Relating these two categories,

212


however, requires redefining the role of internal causation in the account of the causative

alternation. The verbs which are classified as anticausative in our experiments do alter-

nate in English, while internal causation has been used to explain why some verbs do

not alternate in English (see Section 5.2.1). Anticausative verbs include both the verbs

which do and which do not alternate in English, but all of these verbs can be expected

to alternate in fewer languages than the verbs in the middle of the scale.

The question of whether there is a difference between the class of cause-unspecified

and causative verbs remains open leading to another potential question. If it turns out

that these two classes cannot be distinguished, this raises the question of why the two

extremes of the scale behave differently with respect to classification. The data collected

in our experiments do not seem to provide enough empirical evidence to address these

issues. Although some tendencies seem to emerge in our classification experiments,

more data from more languages would need to be analysed before some answers to these

questions can be offered.

5.4.2. Cross-linguistic variation in English and German

Unlike the previous research which is either monolingual or typological, we choose to take

a micro-comparative approach and to study the use of lexical causatives in English and

German at the level of token. We consider these two languages, which are genetically

and geographically very close, a minimal pair. We can expect fewer lexical types to

be differently realised in English and German than it would be the case in two distant

languages, with fewer potential sources of variation. On the other hand, if a lexical type

is inconsistently used in English and German, inconsistent realisations of the type can

be expected in any two languages. This approach is in line with some recent trends in

theoretical linguistics (discussed in Section 3.1.1 in Chapter 3).

Despite the fact that English and German are closely related languages, systematically

different realisations of lexical causatives could be expected on the basis of the grammat-

ical and lexical differences that have already been identified in the literature. It has been

noticed that the sets of alternating verbs in these languages are not the same (Schafer

2009). English verbs of moving, such as run, swim, walk, fly alternate, having both

213


anticausative and causative version. Their lexical counterparts in German can only be

found as intransitive. The causative use of these verbs in English necessarily requires a

transformation in the German translation. On the other hand, some verbs can alternate

in German, but not in English. For example, the verb verstarken (’reinforce’) in German

can have the anticausative version (sich verstarken), while in English, only the causative

version is possible. The equivalent of the German anticausative verb is the expression

become strong. At a more general level, it has been claimed that the relations between

the elements of the argument structure of German verbs are more specified than those

in English verbs, especially in prefixed verbs (Hawkins 1986). Given the opinions that

the degree of specificity of verbs’ meaning can influence the alternation, which have been

put forward in the qualitative analyses of lexical causatives (see Section 5.2 for more

details), this difference might have influence on the way the verbs are realised in the two

languages. From the morphological point of view, the alternation is differently marked

in the two languages. While English shows preference for labile verb pairs (see Section

5.5), German uses both anticausative marking and labil pairs. This difference is another

factor than could potentially influence the realisations of the verbs.

Although we have not examined the influence of these factors directly, the results of our

experiments suggest that including the data from German in the automatic estimation of

the likelihood of external causation changes the estimation so that it corresponds better

to the typological ranking of verbs in the three-way classification setting. This can be

interpreted as an indication that lexical causatives are realised differently in English

and German, despite the fact that the two languages are similar in many respects. The

difference is big enough to neutralise some language-specific trends in the realisations of

lexical causatives, such as, for example, the preference for transitive clauses in English,

discussed in the beginning of Section 5.3.4.


Studying the alternation in lexical causatives is not only interesting for theoretical lin-

guistics. As discussed in Chapter 2, formal representation of the meaning of verbs are

extensively used in natural language processing too. Analysing the predicate-argument

214

5.5. Related work

structure of verbs proves important for tasks such as word sense disambiguation (La-

pata and Brew 2004), semantic role labelling (Marquez et al. 2008), cross-linguistic

transfer of semantic annotation (Pado and Lapata 2009; Fung et al. 2007; van der Plas

et al. 2011). Several large-scale projects have been undertaken to represent semantic

properties of verbs explicitly in lexicons such as Word Net (Fellbaum 1998), Verb Net

(Kipper Schuler 2005), and PropBank Palmer et al. (2005a).

Since the causative alternation involves most of verbs, identifying the properties of verbs

which allow them to alternate is important for developing representations of the meaning

of verbs in general. The findings of our experiments provide new facts which could be

useful in two natural language processing domains. First, the position of a verb on the

scale of the likelihood of external causation can be used to predict the likelihood for a

clause to be transformed across languages. Generally, the verbs which are in the middle

of the scale can be expected to be used in a parallel fashion across languages, while

placement of a verb on the extremes of the scale gives rise to divergent realisations.

However, studying other factors involved in the realisations in each particular language

would be required to predict the exact transformation. Second, the knowledge about

the likelihood of external causation might be helpful in the task of detecting implicit

arguments of verbs and, especially deverbal nouns (Gerber and Chai 2012; Roth and

Frank 2012). Knowing, for example, that a verb is on the causative side of the scale

increases the probability of an implicit causer if an explicit causer is not detected in a

particular instance of the verb.

5.5. Related work

Frequency distributions of transitive and intransitive realisations of lexical causatives in

language corpora have been extensively studied in natural language processing, start-

ing with the work on verb subcategorisation (Briscoe and Carroll 1997) and argument

alternations (McCarthy and Korhonen 1998; Lapata 1999) to current general distribu-

tional approaches to the meaning of words (Baroni and Lenci 2010) (see Section 2.3 in

Chapter 2 for more details). Here we focus on the work which addresses the notion of

external causation itself in a theoretical framework.

215


McKoon and Macfarland (2000) address the distinction between verbs denoting inter-

nally caused and externally caused events. Their corpus study of twenty-one verb defined

in the linguistic literature as internally caused change-of-state verbs and fourteen verbs

defined as externally caused change-of-state verbs, show that the appearance of these

verbs as causative (transitive) and anticausative (intransitive) cannot be used as a di-

agnostic for the kind of meaning that they are attributed.

Since internally caused change-of-state verbs do not enter the alternation, they were

expected to be found in intransitive clauses only. This, however, was not the case. The

probability for some of these verbs to occur in a transitive clause is actually quite high

(0.63 for the verb corrode, for example). More importantly, no difference was found in

the probability of the verbs denoting internally caused and externally caused events to

occur as transitive or as intransitive. This means that the acceptability judgements used

in the qualitative analysis do not apply to all the verbs in question, and, also, not to all

the instances of these verbs.

Even though the most obvious prediction concerning the corpus instances of the two

groups of verbs was not confirmed, the corpus data were still found to support the dis-

tinction between the two groups. Examining 50 randomly selected instances of transitive

uses of each of the studied verbs, McKoon and Macfarland (2000) find that, when used in

a transitive clause, internally caused change-of-state verbs tend to occur with a limited

set of subjects, while externally caused verbs can occur with a wider range of subjects.

This difference is statistically significant.

The relation between frequencies of certain uses and the lexical semantics of English

verbs is explored by Merlo and Stevenson (2001) in the context of automatic verb clas-

sification. Merlo and Stevenson (2001) show that information collected from instances

of verbs in a corpus can be used to distinguish between three different classes which

all include verbs that alternate between transitive and intransitive use. The classes in

question are manner of motion verbs (5.20), which alternate only in a limited number of

languages, externally caused change of state verbs (5.21), alternating across languages,

and performance/creation verbs, which are not lexical causatives (5.22).

(5.20) a. The horse raced past the barn.

b. The jockey raced the horse past the barn.

216

5.5. Related work

(5.21) a. The butter melted in the pan.

b. The cook melted the butter in the pan.

(5.22) a. The boy played.

b. The boy played soccer.

In the classification task, the verbs are described in terms of features that quantify the

relevant aspects of verbs’ use on the basis of corpus data. The three main features

are derived from the linguistic analysis of the verbs’ argument structure. The feature

transitivity is used to capture the fact that transitive use is not equally common for all

the verbs. It is very uncommon for manner of motion verbs (5.20b), much more common

for change of state verbs (5.21b), and, finally, very common for performance/creation

verbs (5.22b). This means that manner of motion verbs are expected to have consistently

a low value for this feature. Change of state verbs are expected to have a middle value

for this feature, while a high value of this feature is expected for performance/creation

verbs. The feature causativity represents the fact that, in the causative alternation, the

same lexical items can occur both as subjects and as objects of the same verb. This

feature is expected to distinguish between the two causative classes and the performance

class. The feature animacy is used to distinguish between the verbs that tend to have

animate subjects (manner of motion and performance verbs) and those that do not

(change of state verbs).

The results of the study show that the classifier performs best if all the features are

used. They also show that the discriminative value of the features differs when they are

used separately and when they are used together, which means that information about

the use of verbs that they encode is partially overlapping.

In our study, we draw on the fact that the lexical properties of verbs are reflected in

the way they are used in a corpus, established by the presented empirical approaches to

the causative alternation. As in these studies, we consider frequencies of certain uses of

verbs an observable and measurable property which serves as empirical evidence of the

lexical properties. We explore this relationship further, relating it to a deeper level of

theoretical semantic analysis of verbs and to the typological distribution of grammatical

features.

217



The experiments presented in this chapter provide empirical evidence that contribute

to better understanding of the relationship between the semantics of lexical causatives,

their formal morphological and syntactic properties, and the variation in their use. First,

we have shown that the distribution of morphological marking on lexical causatives

across a wide range of languages correlates with the distribution of their two syntactic

realisations (transitive and intransitive) in a corpus of a single language. We have argued

that the underlying cause of this correlation can be the meaning of lexical causatives,

more precisely, the likelihood of an external causer in an event described by a verb.

We have then proposed a monolingual corpus-based measure of the likelihood of an

external causer which is automatically calculated for a wide range of alternating verbs.

Having assigned the likelihood values to a large sample of verbs, we have analysed the

distribution of the syntactic alternants of the verbs in the cross-linguistic English and

German realisations. The analysis was performed on data automatically extracted from

parallel corpora. It showed that the likelihood of external causation in the meaning of

verbs influences the cross-linguistic variation. The realisations which are consistent with

this value tend to be realised across languages in a parallel fashion. Contrary to this,

the realisations in one language which are not consistent with the meaning of the verb

tend to be transformed into consistent realisations in the other language.

We have shown that automatic assessment of the likelihood of external causation in

the meaning of verbs is more congruent with the typological data if it is based on the

information on the realisations of lexical causatives in two languages than if it ist based

on the monolingual information, despite the fact that the two studied languages, English

and German, are typologically and geographically very close. To demonstrated this, we

have designed a probabilistic model which classifies verbs into external causation classes

taking into account cross-linguistic data. We have then evaluated the classification

against verb ordering based on the typological distribution of causative and anticausative

morphological marking of lexical causatives, as well as with the ordering based on the

distribution of the syntactic alternants of lexical causatives in a monolingual corpus.

To address the ongoing theoretical discussion on the number of semantic classes of lexical

218


causatives, we have tested two versions of the model: a two-class and a three-class

version. These tests have not provided conclusive results, but they have pointed out

some potential tendencies which can be further investigated.

219

6. Unlexicalised learning of event

duration using parallel corpora

6.1. Introduction

Sentences denote events and states that can last from less than a second to an unlimited

time span. The time span in which an event or a state takes place is understood mostly

implicitly. The interpretation of the time in which the event takes place is sometimes

influenced by the adverbials found in the sentence or by other structural elements, but the

time span itself is hardly ever explicitly formulated. Consider the following example:

(6.1) Winston stayed in the shop for two hours.

The time span for which the relation expressed in (6.1) holds is specified with the

adverbial two hours. However, even in this sentence, where the time adverbial is as

explicit as it can be in natural language, we understand that Winston stayed in the shop

some time “around two hours and not more” due to the automatic inference mechanisms

called conversational implicatures (Grice 1975; Levinson 1983; Moeschler and Reboul

1994). The time span in this sentence is clearly not meant to be interpreted as “at least

two hours”, which is, in fact, the truth-conditional meaning of the adverbial. Eliminating

this second interpretation is based on implicit understanding.

Sentences of natural language provide various clues to infer the implied time interpreta-

tion. We illustrate how the clues guide our interpretation with the examples in (6.2).

(6.2) a. Winston looked around himself for a while before he quickly put the book in

the bag.

221

6. Unlexicalised learning of event duration using parallel corpora

b. Winston looked around himself quickly before he started putting the book

in the bag.

c. Winston looked around himself quickly before he started putting the books

in the bag.

The adverbials for a while and quickly in (6.2a) do not directly quantify the duration

of the events of looking around and putting the book in the bag respectively, but rather

suggest appropriate readings. For instance, it is appropriate to understand that the

former event lasted longer than the latter and that both of them did not last more

than a minute. If we rearrange the constituents as in (6.2b), then the appropriate

understanding is that the former event was shorter than the latter. The event of putting

the book in the bag is obviously longer in (6.2b) than in (6.2a), but its duration is also

less specified. The event of putting the books in the bag in (6.2c) is interpreted as the

longest of the three, which is due to the plural form of the head of the direct object

(books) (See Krifka (1998) for more details on the relationship between the meaning of

the direct object of a verb and the temporal properties of the event described by the

verb.)

The examples in (6.2) show that the interpretation of the duration of an event depends

not just on the time adverbials, but also on the semantics of the verbs, and even on the

semantics of their complements. All the events in (6.2) are interpreted as lasting up to

several minutes, with the described variation. The common time range can be related to

the meaning of the verbs which are used to describe the events. The time span is more

specified with the time adverbials that are used in the sentences, with each adverbial

setting resulting in a slightly different interpretation, as in (6.2a-b). Finally, as we see

in (6.2c), a particular interpretation can be the result of a verb-object combination.

While time adverbials can guide our intuition in selecting the most appropriate time

interpretation, they cannot be taken as reliable quantifiers of events. The time span

over which an event holds can be underspecified, despite a very precise time expressed

by the time adverbial. This is the case in (6.3a).

(6.3) a. On 29th November 1984 at 7:55, Winston turned the switch.

222

6.1. Introduction

b. On 29th November 1984 at 7:55, Winston turned the switch and the whole

building disappeared.

The default interpretation of (6.3a) is that the event is very short, shorter than a minute,

and that it takes place at some time around 7:55. This interpretation is suggested in

the context given in (6.3b). Note that the causal relation between Winston turning the

switch and the building disappearing is inferred rather than encoded (Lascarides and

Asher 1993; Wilson and Sperber 1998). The same sentence can denote a situation

where the two events are unrelated. Nevertheless, the causal relation is automatically

inferred as a conversational implicature, which then suggests that everything, including

the turning of the switch, happened in a very short time span.

Without additional contextual elements, the sentence in (6.3a) can be assigned another

interpretation. It can denote a situation where Winston turned the switch over an

unlimited time span which includes the point described with the time adverbial. We do

not know when the event started, we know that it is true on 29th November 1984 at

7:55, and we do not know whether (and when) it ended.

The essentially implicit nature of understanding how long an event lasted is what makes

the task of automatic identification of event duration especially difficult. In automatic

processing of natural language, the intuition used by a human interpreter to detect the

intended implicit meaning has to be replaced by more explicit and rational reasoning in

which the relationship between the linguistic clues and the time interpretation has to be

fully specified. Yet the information about the duration of an event can rarely be read

directly from the lexical representation of the time adverbials in the sentences, as the

examples listed above show. Some reasoning is required to assign the correct intended

interpretation even to time adverbials such as two hours in (6.1), which appear explicit

and straightforward. The relationship between the linguistic clues and the resulting

time interpretation is even less specified in (6.2), and it is unspecified in (6.3a), where

it cannot be used for narrowing down the possible interpretations. Apart from the fact

that one of the two interpretations is generally more likely, the sentence (6.3a) does not

contain any elements which can point to the interpretation which is appropriate in the

given contexts. Even though there is a time adverbial in the sentence, it is not useful

223


for disambiguating between the two interpretations. The clues for the disambiguation

would have to be found elsewhere.

The work on automatic identification of event duration, despite the difficulty, is moti-

vated by the importance that it has in natural language understanding. Using natural

language sentences as statements on the basis of which we want to infer some actual

state of affairs often requires knowing the time span over which a certain situation or

event is true. For instance, the sentence in (6.3a) can be taken as a basis for inferring

the state of the switch at 8:00, or whether Winston still turned the switch at 8:30. The

interpretation suggested in (6.3b) makes it almost certain that the switch is in a different

position at 8:00 than it was before 7:55, and also that the event of Winston turning the

switch is not true at 8:30. These inferences cannot be made starting with the temporally

unbounded interpretation of the same sentence.

One approach to dealing with the incomplete information about the time value of a

natural language sentence is to rely on the semantic representation of the whole discourse

to which the sentence in question belongs. The interpretation of the time expressed in

the sentence is deduced from the representation of the time adverbials (or, possibly,

other linguistic units which can point to the relevant time) found in other sentences,

and from certain knowledge about the structure of the discourse. A narrative discourse

would impose chronological sequencing of sentences, while the sentence sequencing in

an argumentative discourse would be only partly chronological, with many inversions

(stepping back in time). Knowing how the events are sequenced is helpful in determining

which one is true at which point in time. There are, however, two problems with this

approach. First, it is computationally complex, because it requires working with complex

representations of big chunks of discourse at the same time. Second, pieces of discourse

usually do not belong to a single type, but rather include the characteristics of several

types at the same time (Biber 1991). With multiple types present in the same sequence

of sentences, it is hard to see which type should be used for which sentence.

Another approach, the one which we pursue in this study, is to rely on an elaborate

semantic representation of verbal predicates, which are at the core of event denotation.

In this study, we search for the elements of the lexical representation of verbs which can

provide useful information for determining the duration of the events that they describe.

224

6.1. Introduction

We make hypotheses about the nature of these elements on the basis of theoretical

arguments put forward in the linguistic literature. More specifically, we study verb

aspect as a lexical and grammatical category that has been widely discussed in relation

with temporal properties of the meaning of verbs. We consider a range of theoretical

insights about what semantic traits are encoded by verb aspect and how these traits can

be related to event duration, which is a category of interest for automatic construction

of real-world representations in natural language understanding.

In the experiment presented in this study, we explore linguistic encoding and the pos-

sibility of cross-linguistic transfer of aspect information. It is a well-known fact that

Slavic languages encode verb aspect in a much more consistent way than most of the

other European languages. However, the mechanism which is used to encode aspect in

Slavic languages is lexical derivation of verbs, not syntactic or morphological rules. The

consequence of the lexical nature of aspect encoding is that the derivational patterns

are not regular, but rather idiosyncratic and unpredictable, presenting numerous chal-

lenges for generalisations. In our study, we take a data-driven probabilistic approach to

aspect encoding in Serbian as a Slavic language. We develop a cross-linguistic aspectual

representation of verbs using automatic token-level alignment of verbs in English and

Serbian. We then apply this representation to model and predict event duration classes.

This approach is based on an assumption that aspectual meaning is preserved across

languages at the level of lexical items (see Section 3.1 in Chapter 3 for a more elaborate

discussion).

In the remaining of this chapter, we first explain the theoretical notions on which our

experiments are based. In Section 6.2.1, we present aspectual classes of verbs most

commonly referred to in the literature and describe their potential relationship with

event duration. The observable properties of semantic aspectual classification of verbs

in general are discussed in Section 6.2.2. Since, unlike English, Serbian encodes temporal

properties of verb meaning in a relatively consistent way, we then show how aspectual

information is encoded in the morphology of Serbian verbs in Section 6.2.3. After laying

down the theoretical foundations, we describe our cross-linguistic data driven approach

to representing verb aspect (Section 6.3). We proceed by describing the experiment

performed to test whether event duration is predictable from the representation of aspect

in Section 6.4. Our statistical model which is designed to learn event duration from the

225


representation of aspect is described in Section 6.4.1. In Section 6.4.2, we describe the

data used for the experiment, as well as various experimental settings. The results of

the evaluation are discussed in Section 6.4.2. After general discussion (Section 6.5), our

approach is compared to related work in Section 6.6.


Events described by verbs have different temporal properties: they can start and end

at some point in time, they can be long, short, instantaneous, or somewhere in between

these categories, they can overlap with other events or a particular point in time, the

overlap can be partial or full, and so on. However, not all of these properties are

equally relevant for the grammatical meaning of verb. Some temporal properties give

rise to certain grammatical structures, while others are grammatically irrelevant. The

main issue in linguistic theory of verb aspect is the kind of temporal meaning which it

encodes. In this section, we give an overview of the main accounts of the relationship

between the temporal meaning and the form of verbs.

6.2.1. Aspectual classes of verbs

It has been argued in the linguistic literature that verbs can be divided into a (small)

number of aspectual classes. A verb’s membership in an aspectual class is argued to play

an important role in interpreting time relations in discourse. Dowty (1986) discusses

contrasts such as the one shown in the sentences in (6.4a-b).

(6.4) a. John entered the president’s office. The president woke up.

b. John entered the president’s office. The clock on the wall ticked loudly.

The interpretation of (6.4a) is that the president woke up after John entered his office,

while the interpretation of (6.4b) is that the clock ticked loudly before and at the same

time with John’s entering the president’s office. Dowty (1986) argues that the contrast

226


Eventuality described by a verb

Unbounded Bounded

Statehave,know,believe

Activityswim,walk,talk

Achievementarrive,realise,learn

Accomplishmentgive,teach,paint

Figure 6.1.: Traditional lexical verb aspect classes, known as Vendler’s classes

is due to the fact that the verbs wake up and tick belong to two different aspectual

classes.

The aspectual classification of verbs depends on a set of lexical properties which describe

the dynamics and the duration of the event which is described by the verb. A verb

can describe a stative relation (as in (6.1)). We say that these verbs describe states.

Otherwise, a verb can describe a dynamic action (as in (6.2-6.4)). States are usually

considered temporally unbounded, while actions can be unbounded and bounded.1

The temporal boundaries to which we refer here are those that are implicit to the

meaning of the verb. Although the state in (6.1), for example, is temporally bounded

to two hours, this is imposed by the time adverbial. The meaning of the verb stay itself

does not imply that there is a start or an end point of the state described by it. In

contrast to this, the meaning of a verb such as wake up in (6.4a) does imply that there

is a point in time where the action described by the verb (waking up in this case) is

completed. Such verbs are said to describe temporally bounded events, usually termed

as telic actions in the literature. (In this sense, states are always atelic.) Actions can

1The term event, as it is used in the linguistic literature, sometimes does not include states, butonly dynamic aspectual types. The general term which includes all aspectual types is eventuality.However, this distinction is not always made, especially not in computational approaches, and theterm event is used in the more general sense, the one covering all aspectual types. In our study, weuse the term event in the general sense.

227


also be temporally unbounded (atelic). This is the case with the clock ticking in (6.4b).

Even though this action consists of repeated, temporally bounded actions of ticking, the

verb is understood as describing a single temporally unbounded action, usually termed

as an activity. The difference in the existence of an implicit time boundary in the

interpretation of the verbs wake up and tick is precisely what creates the difference in

the interpretation of the event ordering in (6.4a) and (6.4b).

Temporally bounded actions can be bounded in different ways. Most commonly a dis-

tinction is made between the actions that last for some time before they finish, known

as accomplishments and the actions which both begin and end in a single point of time,

known as achievements. Typical examples of verbs that describe accomplishments are

build, give, teach, paint, and those that describe achievements are arrive, realise, learn.

Accomplishments are usually thought of as telic actions, as they point to the end of an

action. Achievements, on the other hand, are frequently described as inchoative, which

means that they point to the beginning of an action.

This taxonomy of four aspectual types, summarised in Fig 6.1, is often referred to as

Vendler’s aspectual classes (Vendler 1967). It has a long tradition in the linguistic

theory, but it cannot be taken as a reference classification, as more recent work on verb

aspect shows.

The distinction between the entities which are temporally unbounded and those which

are bounded seems much easier to make than the distinctions referred to at the sec-

ond level of the classification. Dowty (1986) proposes a precise semantic criterion for

distinguishing between the two:

(a) A sentence2 ϕ is stative iff it follows from the truth of ϕ at an interval I

that ϕ is true at all subintervals of I. (e.g. if John was asleep from 1:00

until 2:00PM, then he was asleep at all subintervals of this interval: be

asleep is a stative).

(b) A sentence ϕ is an activity (or energeia) iff it follows from the truth

of ϕ at an interval I that ϕ is true of all subintervals of I down to a

2Although the units discussed here are sentences, Dowty (1986) explicitly applies the same criteria tolexical items and functional categories.

228


certain limit in size (e.g. if John walked from 1:00 until 2:00 PM, then

most subintervals of this time are times at which John walked; walk is

an activity.)

(c) A sentence ϕ is an accomplishment/achievement (or kinesis) iff it follows

from the truth of ϕ at an interval I that ϕ is false at all subintervals of

I. (E.g. if John built a house in exactly the interval from September 1

until June 1, then it is false that he built a house in any subinterval of

this interval: build a house is an accomplishment/achievement.)

(Dowty 1986: p. 42)

There are two points to note about this criterion. First, the main difference is made

between (a) and (b) on one side and (c) on the other. The difference between (a)

and (b) is only in the degree to which the implication is true: in (a) it is true for all

subintervals, while in (b), it is true for most subintervals. This difference does not matter

in the contrasts illustrated in (6.4). With respect to the interpretation of time ordering,

items defined in (a) and in (b) behave in the same way. Second, the distinction between

accomplishments and achievements is not made at all. The reason for this is not just

the fact that the distinction, like the distinction between (a) and (b), does not play a

role in the contrasts addressed in Dowty’s study, but also the fact that there are no

clear criteria for distinguishing between the two. Dowty (1986) argues that the duration

criterion evoked in the literature does not apply.

Relating the distinction between bounded and unbounded entities to only dynamic types,

as we have been doing so far for clarity of presentation of the traditional taxonomy, does

not entirely correspond to the real semantics of verbal predicates. Marın and McNally

(2011) show that some verbs which would traditionally be classified as achievements

(Spanish aburrirse ‘to be/become bored’ and enfadarse ‘to become angry’) are states,

in fact, even though they are temporally bounded (inchoative). Other authors have

proposed other criteria for defining aspectual classes. Most approaches analyse events

in an attempt to define their components, such as start, or result. The classes are

then derived as different combination of the components (Pustejovsky 1995; Rothstein

2008; Ramchand 2008).

229


In our study, no particular classification or event structure is adopted. We use the notions

of temporal boundedness and duration, which are well defined and endorsed by most of

the studies discussed above, but we do not adopt any of the traditional classes which are

defined by particular combinations of the values of these categories. Since traditional

aspectual classes are questionable, as shown in the discussion above, we propose our own

approach to combining the values of boundedness and duration in forming aspectual

classes. Our representation of aspect is based on the traditionally discussed notions, but

the resulting categories do not correspond to any of the traditionally defined clsasses.

As opposed to Dowty (1986), our study does not address orderings of events. We are

interested in the duration of each event separately. Regarding the examples in (6.4),

for instance, we are not interested in knowing whether the president woke up before or

after John entered his office. Our questions are: How long does John’s entering the

president’s office last? How long is the president’s waking up and the clock’s ticking?

They are related to the sequencing of the parts of discourse, but they can be treated

separately.

We try to answer these questions using the knowledge about verbs’ aspectual classes,

which are defined based on the notion of temporal boundedness. The intuition behind

this goal is that temporal boundedness and duration are related. It is reasonable to

expect that short events are temporally bounded. It is easier to imagine a time boundary

in something that lasts several seconds than in something that lasts a hundred years.

Long events can be expected to be less temporally bounded. Note that our expectations

are probabilistic. We do not exclude the possibility for a short event to be temporally

unbounded and for a long event to be bounded. However, we expect short temporally

unbounded events to be less likely than than short temporally bounded events and long

temporally bounded events to be less likely than long temporally unbounded event. We

expect these dependencies to be strong enough so that the duration of an event can be

predicted from aspectual properties of the verb that expresses it.

230


Tense Aspectual interpretation

Present/Past Simple unspecifiedPresent/Past Continuous → ActivitiesPresent/Past Perfect → Bounded (achievements/accomplishments)

Table 6.1.: A relationship between English verb tenses and aspectual classes.

6.2.2. Observable traits of verb aspect

The criterion for distinguishing temporally bounded and unbounded verbs defined by

Dowty (1986) is a truth-conditional test which is well suited for querying human logical

intuition. To perform such a test automatically, a system would need a comprehensive

knowledge database, with all truth-conditions and inference rules explicitly encoded for

each expression. The size of such a database and the resources needed to create it, as

well as to search it, are hard to asses, but such a project would certainly be a challenging

one. What would be more suitable for automatic identification of temporal boundedness

is to be able to observe formal differences between unbounded and bounded events. This

brings up the question of how aspectual classes can be observed in verbs.

The form of the verbs listed in Fig. 6.1, for example, clearly does not vary depending

on the class membership: verbs belonging to the same class have nothing in common.

By considering only the form of a verb, we cannot determine its aspectual class.

When verbs are used in a sentence, however, they receive a tense form, and some of the

verb tenses in English do encode certain aspectual properties. As illustrated in Table

6.1, continuous tenses tend to refer to activities, while perfect tenses indicate that the

event is temporally bounded.

The tense with which a verb is used can override the inherent aspectual class of the verb

lexical entry. For example, the verb realise is usually classified as an achievement, but

the Present Continuous Tense form as the one in (6.5), cannot be assigned this class.

(6.5) People are slowly realising what is going on.

231


The tense form of English verbs is a useful marker for identifying their aspect, but the

range of the examples in which this relation can be used for automatic identification

is limited to the sentences which contain either continuous or perfect tense. However,

sentences with marked continuous or perfect tense are far less frequent than sentences

with simple or bare forms, which do not point to any particular aspectual class.

Another potential source of observations are distributional properties of verbs. A number

of distributional tests to identify aspectual classes have been proposed in the literature,

starting with Dowty (1979). Interestingly, the proposed tests do not make reference

to the distinction between bounded and unbounded events, but to the second level

of the traditional taxonomy (Fig. 6.1). For example, the most famous test of com-

patibility with in-adverbials vs. for-adverbials, shown in (6.6), differentiates between

states and activities on one side and accomplishments on the other. States and activites

are compatible with the for-adverbials, and accomplishments are compatible with the

in-adverbials. This test does not apply to achievements. Other tests will distinguish

between, for example, states and other classes and so on.

(6.6) a. State:

Winston stayed in the shop for two hours / ??in two hours.

b. Activity :

The clock in the presidents office ticked for two hours / ??in two hours.

c. Accomplishment :

Winston put the books in the bag ??for two seconds / in two seconds.

c. Achievement :

The president woke up ??for two seconds / ??in two seconds.

Apart from the fact that the categories which are identified with the distributional tests

are not clearly defined, the problem with such tests is that English verbs are highly

ambiguous between different aspectual readings so that adverbials can impose almost

any given reading. The use of the verbs in (6.6) with the compatible adverbials is

preferred, but the use with the incompatible adverbials is not ungrammatical. The

incompatible adverbial imposes a less expected, potentially marginal reading, but, with

232


this reading, the sentence remains grammatical. Moreover, many verbs are perfectly

compatible with different contexts, such as write in (6.7).

(6.7) a. Activity:

Winston wrote the message for two hours.

b. Accomplishment:

Winston wrote the message in two hours.

Unlike English, other languages make formal differences between aspectual classes and

these differences can be observed in the structure of verbal lexical entries. The form of

verbs in Slavic languages, for example, famously depends on some aspectual properties

of the events that they describe. In the remaining of this section, we show how this

marking is realised in Serbian, as one of the Slavic languages.

6.2.3. Aspect encoding in the morphology of Serbian verbs

The inventory of Serbian verbs contains different entries for describing temporally un-

bounded and temporally bounded events. Consider, for example, the Serbian equivalents

of the sentence in (6.7), given in (6.8-6.9). The verbs pisao in (6.8) and napisao in (6.9)

constitute a pair of lexical entries in fully complementary distribution: pisati (infinitive

form of pisao) is used for temporally unbounded events, and napisati is used for tem-

porally bounded events. As we can see in (6.8-6.9), exchanging the forms between the

unbounded and bounded contexts makes the sentences ungrammatical. The forms that

are used in the temporally unbounded context are called imperfective and those that are

used in the temporally bounded context are called perfective.

(6.8) VinstonWinston-nom

jeaux

pisao/*napisaowrote

porukumessage-acc

dvatwo

sata.hours.

Winston wrote a message for two hours.

(6.9) VinstonWinston-nom

jeaux

napisao/*pisaowrote

porukumessage-acc

zafor

dvatwo

sata.hours.

Winston wrote a message in two hours.

233


pref(x) = P suff(pref(x)) = I pref(suff(pref(x))) = P

’complete a specified x’ ’do pref(x) continu-ously or repeatedly’

’complete multiplepref(x)’

skuvati — —’cookP ’

prokuvati prokuvavati isprokuvavati’boilP briefly’

iskuvati iskuvavati iziskuvavati’cookP well’

otkuvati otkuvavati izotkuvavati’cleanP by boiling’

zakuvati zakuvavati izakuvavati’addP something intoboiling liquid’

Table 6.2.: Serbian lexical derivations, e.g. x = kuvati (I) ’cookI ’, basic form; I standsfor imperfective, P for perfective.

The two verbs in (6.8-6.9) are obviously morphologically related. The perfective verb

is derived from the imperfective by adding the prefix na-. This case represents the

simplest and the most straightforward relation between an imperfective and a perfective

verb, usually considered prototypical. In reality, the derivations are far more complex,

involving both lexical and aspectual modifications of the basic entry. The category of

temporal boundedness underlies the two major aspectual classes in Serbian (perfective

and imperfective), but it also interacts with some other factors resulting in a more fine-

grained classification, which does not necessarily match the classifications mentioned in

Section 6.2.1. An illustration of derivations involving the verb kuvati (cook) is given in

Table 6.2.

We can see in Table 6.2 that the verbs are organized into aspectual sequences rather

than pairs. Multiple affixes can be added to the same basic verb, modifying its meaning

and its aspect at different levels. Each column in the table represents a step in the

derivation. Each step can be seen as a function that applies to the result of the previous

step. The forms in the first column are the result of adding a prefix to the basic form.

The basic form is imperfective and adding the prefix turns it into a perfective. This

234


derivation is in many ways similar to the attachment of particles to English verbs, as

the translations of the prefixed forms suggest (Milicevic 2004). Adding a prefix results

in a more specified meaning of the basic verb by introducing an additional resultative

predication into the verb’s lexical representation (Arsenijevic 2007). We say that this

derivation indirectly encodes verb aspect because prefixes are not aspectual morphemes.

The change of the aspect is a consequence of the fact that the result state introduced

by the prefix makes the event temporally bounded.

In some cases, this derivation can be further modified as shown in the second column. By

attaching a suffix, the verb receives a new imperfective interpretation which is ambiguous

between progressive and iterative meaning, similar to the interpretation of tick in (6.4b).

(Historically, the suffix is iterative.) This new imperfective, sometimes referred to as

secondary imperfective, is necessarily different than the starting imperfective of the basic

form. The forms in the second column express events containing the resultative predicate

introduced by the prefix, but with the time boundary suppressed by the suffix.

Finally, the forms in the third column are again perfective, but this perfective is different

from the one in the first column. These forms can be regarded as describing plural

bounded events. In actual language use, these forms are much less frequent than the

others. They can be found only in big samples of text, which is why we do not consider

them in our experiments.

Examples in (6.10-6.13) illustrate typical uses of the described forms of Serbian verbs.

(6.10) Basic imperfective:

VinstonWinston-nom

jeaux

cestooften

kuvao.cooked.

Winston often cooked.

(6.11) Prefixed perfective:

VinstonWinston-nom

jeaux

prokuvaoboiled

casuglass-acc

vode.water-gen.

Winston boiled a glass of water.

(6.12) Secondary imperfective:

235


VinstonWinston-nom

jeaux

prokuvavaoboiled

casuglass-acc

vodewater-gen

(kada(when

jeaux

cuoheard

glas).sound-acc).

Winston was boiling a glass of water when he heard the sound.

(6.13) Double-prefix plural perfective:

VinstonWinston-nom

jeaux

isprokuvavaoboiled

sveall-acc

caseglasses-acc

vode.water-gen.

Winston boiled all the glasses of water.

Not all prefixed verbs can be further modified. The verb skuvati in Table 6.2, for exam-

ple, does not have the forms which would belong to the second and the third column.

This phenomenon has been widely discussed in the literature, with many authors trying

to determine what exactly blocks further derivations. It has been argued that prefixes

can be divided into lexical (or inner) and superlexical (or outer) and that further deriva-

tions are not possible if the prefix is superlexical. This account, however, remains subject

of debate (Svenonius 2004b; Arsenijevic 2006; Zaucer 2010). We ignore this difference

considering that there are no structural differences between the prefixes.

Lexical and aspectual derivations are also possible with verbs whose basic form has

perfective meaning. An illustration of this paradigm is given in Table 6.3. The aspect

does not change for these verbs when they are prefixed. The verbs in the first column

are perfective just like the basic form. The rest of derivations proceed in the same way

as for the imperfective basic forms.

There are other patterns of lexical expression of aspectual classes in Serbian. For in-

stance, some verbs are attached a perfective suffix (rather than a prefix) directly to

the basic form (usually used to express very short, semelfactive events), some verbs do

not have the basic form, they are found only with prefixes, some perfective verbs have

no imperfective counterparts and vice versa, etc. However, the examples listed in this

section form a general picture of how aspectual classes are morphologically marked in

Serbian. The summary of possible verb forms is given in Fig. 6.2.

What is important for our study is the fact that aspectual classes are observable in the

verb forms, although the relationship between the form and the meaning is not simple.

236


suff(x) = I ’do x continuously or repeatedly’ → bacati

pref(x) = P suff(pref(x)) = I pref(suff(pref(x))) = P

’complete a specified x’ ’do pref(x) continu-ously or repeatedly’

’complete multiplepref(x)’

prebaciti prebacivati isprebacivatitransferPizbaciti izbacivati izizbacivatithrowP out

ubaciti ubacivati izubacivatithrowP in

odbaciti odbacivati izodbacvatirejectP

Table 6.3.: Serbian lexical derivations, e.g. x = baciti (P) ’throwI ’, basic form; I standsfor imperfective, P for perfective

verb

<outer prefix> <inner prefix> <stem> <suffix> <inflection>

imperfective’-va’’-ja’other

perfective’-nu’

’iz’-’na-’. . .

’iz’-’na-’’u-’’ od-’. . .

. . . tensemood

Figure 6.2.: Serbian verb structure summary

237


Serbian verb morphology encodes aspect only indirectly, but, unlike with English verb

tenses, some kind of aspect information is present in almost all verb uses. Morphological

expression of aspect in Serbian is also potentially less ambiguous, hence more helpful in

determining verb aspect than time adverbials and other elements which can be found in

the context in which the verb is used.

The described derivations can potentially encode numerous aspectual classes. Minimally,

the verbs are divided into temporally bounded (perfective) and temporally unbounded

(imperfective). However, combinations of different stems, prefixes and suffixes can result

in more fine-grained classes. For example, the secondary imperfectives (the third column

in Table 6.2) do not have the same temporal properties as the starting, basic imperfec-

tive. As discussed above, the secondary imperfective contains the resultative meaning

introduced by the prefix, while the bare form does not. As a consequence, the meaning

of the secondary imperfective is more specified and more anchored in the context. This

distinction might prove relevant for event duration. We can expect prefixed imperfective

verbs to describe shorter events than basic imperfective verbs. Further distinctions can

be encoded by potential dependencies between the prefixes and suffixes, or between the

stems and the structural elements.

In our study, we do not explore all the possible encoded meanings, but we do explore

distinctions which are more complex than the simplest distinction between imperfec-

tive and perfective meaning. We represent aspectual classes as combinations of three

attribute-value pairs. The attributes are defined on the basis of the analysis presented

in this section.

6.3. A quantitative representation of aspect based on

cross-linguistic data

Experimental approach to the temporal meaning which we adopt in our study requires

a relatively large set of examples of linguistically expressed events for which both event

duration and verb aspect are known and explicitly encoded. Compiling such a data set

is already a challenging task because we are dealing with intuitive notions referring to

238

6.3. A quantitative representation of aspect based on cross-linguistic data

the phenomena which are hard to observe and measure. Although we all might agree

that some events last longer than others, assessing exact duration of any particular

event is something that we do not normally do. Collecting such assessments for a

large number of cases is a task which requires significant efforts. Collecting verb aspect

assessments is even more complicated because this is a theoretical notion which cannot

be understood without linguistic training. Moreover, assessing an exact verb aspect

value for a particular verb use proves difficult even for trained linguists because there is

no general consensus in the theory on what values this category includes and how they

should be defined.

One of the objectives of our work is, thus, compiling a set of data for our experimental

work. The existing resources provide one part of the information that we need: human

judgments of event duration have been annotated for a set of event instances (Pan et al.

2011). We collect the information about verb aspect for the same set of instances.

We decide to collect our own verb aspect information because of the fact that no theo-

retical account of this category can be taken as standard. Using any existing annotation

necessarily means adopting the view of verb aspect inherent to the annotation. Instead,

we propose our own approach which is data-driven and based on observable structural

elements of verb forms. Since English verb morphology typically does not mark aspect,

we gather the information from the translations of English verbs into Serbian, where,

like in all other Slavic languages, verb aspect is indirectly encoded in the verbs’ mor-

phology. Instead of collecting human judgements, we observe natural language encoding

of verb aspect in Serbian and use these observations to assign verb aspect properties to

the verbs both in Serbian and English. The information used for assignment is automat-

ically extracted from language corpora, which makes this approach especially suitable

for annotating a large number of examples.

We represent aspectual classes of verbs as combinations of values of three attributes.

The first attribute is the binary classification into grammatical perfective and impere-

fective classes. The second and the third attribute (morphological attributes) specify

whether the verb in question is used with a prefix or/and with a suffix respectively. The

combinations of the values of the three attributes represent aspectual classes. We use

these classes to predict event duration, but we do not identify them with any aspectual

239


classes already proposed in the literature. The values for the three attributes are as-

sessed on the basis of the instances of verbs in a parallel corpus, which is described in

the following section.

The values for aspectual attributes of verbs are determined automatically on the basis

of cross-linguisitc realisations of verbs in English and Serbian. Cross-linguistic assign-

ment of aspectual classes is possible due to the fact that aspect is a property of the

meaning of verbs, and the meaning is what is preserved in translating from one lan-

guage to another, while the forms of its expression change. In our case, verb aspect

is morphologically encoded in Serbian verbs, but the same aspectual class can be as-

signed to the corresponding English verbs. In this section, we describe our approach to

cross-linguistic aspectual classification of verbs based on the described representation of

aspect decomposed into attributes.

6.3.1. Corpus and processing

For transferring verb aspect classification from Serbian to English, we need to know

which Serbian verb is the translation of an English verb in a given context. This kind of

information can be obtained from a corpus of translated sentences — a parallel corpus

— which is aligned at the word level, so that the translation is known for each word of

each sentence. For the purpose of predicting event duration on the basis of aspectual

properties of verbs, which is the goal of our experiments, we need a word-aligned parallel

corpus with annotated event duration on one side of the corpus. Such a corpus, however,

does not exist. Examples of annotated event duration are available in English and we

do not have Serbian translations of these sentences. We have to use other resources.

There are several parallel English-Serbian corpora which are currently available (Tiede-

mann 2009; Erjavec 2010). In our current study we only use the Serbian translation

of the novel “1984” by George Orwell, which is created in the MULTEXT-East project

(Krstev et al. 2004; Erjavec 2010). We use this corpus for the convenience of the

manual annotation and literary text genre, which is known to be rich in verbs (Biber

1991). In principle, our methods are applicable to all available parallel corpora.

240


Serbian English

Word Lemma MSD Word Lemma MSD

propagirao propagirati Vmps-sman-n—p proclaiming Vmpp proclaimje jesam Va-p3s-an-y—p his Ds3—sm hissvoju svoj Ps-fsa heresy Ncns heresyjeres jeres Ncfsa–n , # #, # # exulting Vmpp exultzanosio zanositi Vmps-sman-n—p in Sp inse se Q it Pp3ns itnjome ona Pp3fsi

Table 6.4.: An illustration of the MULTEX-East corpus: manually encoded morpholog-ical analysis of a chunk of text in Serbian and its corresponding chunk inEnglish.

The MULTEXT-East parallel corpus is available as an XML database containing two

kinds of manual annotation:

• Morphological annotation — Each word in the text is assigned a lemma and a

code called morphosyntacic definitions (MSD), which is a compact representation

of a number of lexical, morphological, and syntactic categories realised in each

word form. Each category is encoded by a single character in the label. The

first character encodes the part-of-speech (verb, noun, adjective, etc.). The second

character encodes a subclassification for each main category (e.g. main, auxiliary,

modal, copula for verbs, common, proper for nouns etc.). Other characters specify

morphological features that are marked in the word form such as case, number,

tense, voice, mood etc. For example, the MSD label ”Vmps-sman-n—p” denotes

that the word propagirao is a main verb, in the past participle singular mascu-

line active positive form. The last letter indicates that its aspect is imperfective.

The letter “p” in the MSD code stands for “progressive”, but in fact it encodes

“imperfective” as described in Section 6.2.3. An illustration of the morphological

annotation is shown in Table 6.4. Detailed specifications of the labels are provided

in the MULTEXT-East project documentation.

• Sentence alignment — The information about which sentence in English corre-

241


sponds to which sentence in Serbian is provided as an additional layer of annota-

tion.

The corpus is not aligned at the word level. We obtain word alignments, which is the last

piece of information needed for our study, automatically, using the methods described

in the following subsection.

Automatic alignment of English and Serbian verbs in a parallel corpus. We

extract the information about word alignments in the manually aligned Serbian and

English sentences using the system GIZA++ (Och2003), the same tool which is used

in the previous two studies. The input to the system are tokenised sentence-aligned

corpora in the plain text format, with one sentence alignment chunk per line. We use

the XML pointers in the manual alignment file of the MULTEXT-East corpus to convert

the Serbian and English text to the format required by GIZA++. The conversion also

includes removing the morphological annotation temporarily. We then perform word

alignment in both directions: with English as the target language and Serbian as the

target language.

Given the formal definition of word alignment which is used by the system (see Section

3.2 in Chapter 3), the amount and the correctness of word alignment depends very

much on the given direction. It is common practice in machine translation and in other

tasks involving automatic word alignments to use the intersection of both directions

of alignment, that is only the alignments between the words which are aligned in both

directions (Och and Ney 2003; Pado 2007; Bouma et al. 2010; van der Plas et al. 2011).

This approach gives very precise alignments, but only for a relatively small number of

words. We do not follow this approach, since it leaves out many correct alignments

which are potentially useful for our study, as shown in Chapter 4. Instead, we convert

the alignment output into a format suitable for manual inspection. We then manually

compare a sample of the alignments in the two directions and choose the one which gives

more correct alignments. This, in our case, was the alignment with English as the target

language.

Once we have obtained word alignments, we combine them with the morphological anno-

tation from the original corpus to extract only those alignments of English verbs which

are aligned with Serbian verbs. In other words, we only keep alignments between the

242


words which both contain the “Vm” code in their respective morphosyntactic definitions

(see Table 6.4). This simple method does not only select verbs, which we are interested

in in our study, but it also eliminates potentially wrong alignments. If a system aligns

a verb with a noun, or an adjective, or any other category, chances are that this is a

wrong alignment. On the other hand, even if it is possible for a verb in one language

to be aligned with a wrong verb in another language in a sentence which contains more

than one verb, this happens relatively rarely in practice.

6.3.2. Manual aspect classification in Serbian

With word-to-word alignment between English and Serbian verbs and with the manually

annotated aspect code, which is contained in the morphological description of the words

on the Serbian side, we can now see whether English verbs are aligned with perfective or

imperfective Serbian verbs. This will determine the value of the first aspectual attribute

(simple binary aspect classification). We collect the following counts:

• For each verb form on the English side of the corpus:

– the number of times it is aligned with a perfective Serbian verb

– the number of times it is aligned with an imperfective Serbian verb

• For each verb lemma on the English side of the corpus:

– the sum of the alignments of all the forms with a perfective Serbian verb

– the sum of the alignments of all the forms with a perfective Serbian verb

We collect the counts at the level of verb type because some of the verb tenses in English

can indicate a particular aspectual context, as shown in Section 6.2.2. This implies that

aspectual classes assigned to verb forms separately are expected to be more precise than

assigning the same class to all the forms of one lemma.

Summing up the counts for each lemma, on the other hand, is useful as a kind of back-off

for classifying verb forms which are not observed in the parallel corpus. If any other form

243


of the same lemma is observed, then the count of alignments for the lemma is not zero

and the unobserved form can be assigned the value which is assigned to the lemma.

6.3.3. Morphological attributes

The binary grammatical category of perfective and imperfective aspect does not repre-

sent all the aspectual properties of Serbian verbs which are encoded in the morphology.

As shown in Section 6.2.3, these categories interact with other factors and the resulting

morphology encodes a more fine-grained classification. Perfective verbs, for instance, can

be divided into those that are perfective in their basic form (such as baciti in Table 6.3),

those that have become perfective by attaching a prefix (the first column in Table 6.2

and 6.3), or those that are attached a perfective suffix (see Fig. 6.2, these verbs usually

do not have bare forms). Similarly, verbs can be imperfective in their basic form such as

kuvati in Table 6.2, or after they have been attached both a prefix and an imperfective

suffix. We take the presence or the absence of the relevant morphological units in the

structure of Serbian verbs as indicators of different aspectual properties.

For these reasons, in addition to the simple perfectivity, we define two more attributes

for encoding more fine-grained aspectual distinctions of verbs. To collect the counts

needed for this description, we analyse all the Serbian verbs which are identified as

aligned with English verbs in the parallel corpus. We perform the analysis automatically

using our own analyser which implements the rules described in Section 6.2.3. The

structure obtained with our analyser cannot be considered the true structure but rather

an approximation of it. Due to historical changes, some morphemes that are known

to have existed in the structure are not easily recognisable in the present-day forms.

We ignore these elements and treat these verbs as uncompositional. By identifying the

visible morphological segments only, we can still analyse most Serbian verbs.

With the identified segments of the morphological structure of Serbian verbs, we can

now collect the following counts:

• For each verb form on the English side of the corpus:

– the number of times it is aligned with a prefixed Serbian verb

244


– the number of times it is aligned with a suffixed Serbian verb

• For each verb lemma on the English side of the corpus:

– the sum of the alignments of all the forms with a prefixed Serbian verb

– the sum of the alignments of all the forms with a suffixed Serbian verb

Knowing the value of each of the three aspectual attributes of the Serbian alignment of

an English verb form, that is knowing if the Serbian translation of an English verb has

a prefix or not, if it has a suffix or not, and if it is perfective or imperfective, we can

now describe English forms in terms of these attributes, and then use them to predict

the duration of the events expressed by the verbs. We assign to each English verb form

and to each lemma a single value for each of the three aspectual attributes. The values

represent the total of the corpus counts for each type. We explain in the following

subsection how the values are calculated.

6.3.4. Numerical values of aspect attributes

In our cross-linguistic representation, aspect of each English verb form is defined by

a vector of three numbers between 0 and 1. Each number expresses the value of one

attribute. The values are determined based on the observations made in the parallel

corpus. We quantify three aspect attribute in the following way:

• Prefix: This attribute encodes the tendency of English verbs to be word-aligned

with prefixed Serbian verbs. Given the role of prefixes in the derivation of Serbian

verbs, described in Section 6.2.3, such tendency provides two pieces of information

about the event which is described by the English verb:

a) The event is more specified than those that are aligned with Serbian bare

verbs.

b) The event is temporally bounded, unless the verb also tends to be associated

with an imperfective suffix, which can remove the temporal boundary imposed

by the prefix.

245


The value of this attribute is calculated as the proportion of prefixed verbs in the

set of verb alignments for each verb form in English, as shown in (6.14).

Pf(e) =F (sr pref(e))

F (sr(e))(6.14)

where F stands for frequency (= total count in the corpus), e is an English verb,

sr pref(e) is a prefixed Serbian verb aligned with the English verb e, and sr(e) is

any Serbian verb aligned with the English verb e. For example, if an English verb

form is aligned with a Serbian verb in a parallel corpus 9 times, and 7 of these

alignments are Serbian prefixed verbs, the value of Pf = 79

= 0.8.

• Suffix: This attribute encodes the tendency of English verbs to be word-aligned

with Serbian verbs which are attached a suffix. The presence of suffixes in the

Serbian translations of English verbs can mean two opposite things:

a) The event described by the English verb is temporally unbounded, but spec-

ified. This is the case when the suffix is added to derive the secondary imper-

fective in Serbian (the second column in Table 6.2 and 6.3).

b) The event is temporally bounded and very short, which is the case when the

suffix is added to the bare form directly.

The value of this attribute is calculated, similarly to prefix alignments, as the

proportion of verbs containing a suffix in the set of verbs with which an English

verb is aligned, as shown in (6.15).

Sf(e) =F (sr sf(e))

F (sr(e))(6.15)

where e is an English verb, sr suff(e) is a Serbian verb which contains a suffix and

which is aligned with the English verb e, and sr(e) is any Serbian verb aligned with

the English verb e. For example, if the same English verb for which we calculated

the prefix value is aligned with a suffixed Serbian verb 4 times, the value of the

suffix attribute is Sf = 49

= 0.4.

246


• Aspect: This attribute encodes the information extracted from the manually

annotated morphological description of Serbian verbs. It represents the tendency

of an English verb form to be aligned with Serbian verbs tagged as perfective.

This information is especially useful in the case of bare verbs in Serbian, where the

structural information is missing. It is calculated in a similar way as the previous

two values, as shown in (6.16).

Asp(e) =F (sr asp(e))

F (sr(e))(6.16)

where e is an English verb, sr asp(e) is a the perfective aspect annotation of

a Serbian verb aligned with the English verb e, and sr(e) is any Serbian verb

aligned with the English verb e. If the same verb is seen 5 times aligned with

a Serbian perfective verb, the value is calculated as Asp = 59

= 0.5. Note that,

since all the verbs in the corpus are tagged either as perfective or imperfective,

this value determines at the same time the tendency of a verb to be aligned with

imperfective forms in Serbian.

We do not set any threshold to the number of observations that are included in the

measures. We calculate all the three values for all English forms (or lemmas) which are

observed at least once in the parallel corpus. To deal with low frequency items and zero

counts, we apply additive smoothing, which is calculated as in (6.17):

Θi =xi + 1

n+ 2(6.17)

where i ∈ {Pf, Sf,Asp} is one of the aspect attributes, x is the number of observed

alignments of an English verb with a specific value of the attribute, and n is the number

of times the English verb is seen in the parallel corpus. The smoothed values for the

examples used above are ΘPf = 7+19+2

= 0.7, ΘSf = 4+19+2

= 0.4, and ΘAsp = 5+19+2

= 0.5.

We illustrate the resulting aspectual definitions of English verbs with a sample of data

in Table 6.5. The zero values that can be seen in the table are the result of rounding.

247


Verb Prefix Suffix Aspectdeal 0.8 0.5 0.8find 0.9 0.5 0.9owns 0.8 0.8 0.2

crashed 0.2 0.6 0.6thought 0.6 0.0 0.6

hit 0.7 0.1 0.7spent 0.8 0.2 0.3think 0.4 0.1 0.3going 0.4 0.4 0.4

Table 6.5.: A sample of the verb aspect data set.

With the aspectual attributes of English verbs defined using the morphological infor-

mation of Serbian verbs, as we described in this section, we can now perform machine

learning experiments to test whether event duration can be predicted from these de-

scriptions.

6.4. Experiment: Learning event duration with a

statistical model

The main goal of our experiment is to determine if the grammatical notion of verb aspect

encodes the real-world temporal properties of the events. The general intuition behind

our approach, as already discussed in Section 6.2.1, is that the implicit time boundary in

the meaning of a verb and the duration of the event described by it are related. If there is

a time boundary in the lexical representation of a verb, the event described by it is more

likely to be short than if there is no time boundary. Even though time boundaries can

be defined for any event, even for those that last for years, we can expect the boundary

to be implicit to the meaning of only those verbs which describe short events. The

time limit is more prominent in the event whose duration is perceived as limited to a

short time span. This general relationship is then modified in particular cases, such as

248

6.4. Experiment: Learning event duration with a statistical model

English verb tenses (see Section 6.2.2) or secondary imperfectives in Serbian (see Section

6.2.3).

We formalise our hypotheses about the relation between verb aspect and event duration

by means of a statistical model. Representing the aspect attributes with the quantities

based on corpus counts, as described in Section 6.3, is already one part of the model.

The attributes Prefix, Suffix, and Aspect, which we propose, are, in fact, a model of

grammatical aspect. What remains to be specified, in order to construct a full model

of all the notions examined in our research, is how the aspect attributes are related to

event duration, and also how they are related to each other.

The interest for developing such a model is not only practical, but also theoretical. A

model with a sound theoretical background, if successful, is not only expected to make

good predictions improving the performance in the tasks related to natural language

understanding. Such a model, making reference to specific theoretical notions, is also

a means of testing whether these notions actually play a role in the empirical reality.

Being a model of the relationship between the categories in the domain of grammar of

language and those that belong to world knowledge, it can provide new insights into the

functioning of the interface between these two domains.

In the remaining of this section, we first describe the full model which is tested in the

experiment. We then describe the algorithms and the data sets used in the experimental

evaluation, and, in the last subsection, the results of the evaluation.

6.4.1. The model

The full model of the relationship between verb aspect properties and event duration

consists of four variables. The three aspect properties described in Section 6.3 are in-

cluded in the model as separate variables. The fourth variable represents event duration.

In the following list we introduce the notation that we use, summarising the variables,

their values and the expected relationships between them.

T: for Time. This variable represents the information about event duration as assessed

by human annotators. It can take the values “short” and “long”.

249


Pf : for Prefix. This variable encodes the tendency of English verbs to be word-aligned

with prefixed Serbian verbs. It can take the values between 0 and 1, as described

in Section 6.3.

Since the presence of a prefix in Serbian verbs indirectly indicates perfective aspect,

as we show in Section 6.2.3, we expect that higher values of this variable provide

a signal of short event duration, and lower values of long duration.

Sf : for Suffix. This variable encodes the proportion of suffixed verbs in the set of

Serbian verbs with which an English verb form is aligned (Section 6.3).

The expected contribution of this variable to the model is based on the interaction

between verb prefixes and suffixes which is observed in the grammar of Serbian

aspectual derivations. As described in Section 6.2.3, suffixes can be attached to

the verbs which are already attached a prefix, resulting in what is usually called

secondary imperfective. Otherwise, suffixes can be attached to bare forms resulting

in perfective interpretation (see Fig. 6.2). Crossing the values of Pf and Sf is thus

expected to yield a more accurate representation of the temporal properties of

verbs than using any one of the two variables.

Asp: for Aspect. In addition to the formal grammatical elements that indicate the

aspectual class of the Serbian alignments of English verb forms, the model contains

a variable which encodes directly whether the alignments tend to be perfective

or imperfective. Higher values of Asp indicate that the English form tends to

be aligned with perfective Serbian verbs and lower values indicate imperfective

alignments.

The information from this variable is expected to be useful in the cases where

English verb forms are aligned with Serbian verbs which do not bear any formal

marking, but which are still specified for aspect, such as basic forms in Table 6.2

and Table 6.3.

Note that the model does not include any lexical information. We do not use the

information about lexical entries either of English or of Serbian verbs. We also do not

use the form of Serbian prefixes and suffixes, we only observe whether any affix appears

in a verb or not.

250


T

Pf Sf

Asp

Figure 6.3.: Bayesian net model for learning event duration

We formalise the described relationships between the variables in the model by means of

a Bayesian net, shown in Figure 6.3. The general principles of constructing the Bayesian

net model representation are discussed in more detail in Section 3.4.3 in Chapter 3.

As represented by the arrows in Figure 6.3, we assume that Asp and T are conditionally

independent given Pf and Sf. This relationship captures the fact that the information

about verb aspect is important only when the information about verb affixes is not

available. We assume that Sf depends both on T and Pf, which represents the fact that

a suffix can be added for two reasons. First, it can be added as a means of deriving

secondary imperfectives, and this is the case where a prefix is already attached to the

verb. Second, a suffix can be added to a bare form, and, in this cases, it can result

in a perfective. Pf depends only on T, meaning that a prefix is attached only to the

verbs which express events with particular durations (short events). The variable whose

values we predict in the machine learning experiments is T , and the predictors are the

other three variables.

The Bayesian net classifier

We build a supervised classifier which is an implementation of our Bayesian net model

described in Section 6.4.1. Assuming the independence relationships expressed in the

Bayesian net (Figure 6.3), we can decompose the model into smaller factors and calculate

its probability as the product of the probabilities of the factors, as shown in (6.18).

251


P (T, Pf, Sf,Asp) = P (T ) · P (Pf |T ) · P (Sf |Pf, T ) · P (Asp|Pf, Sf) (6.18)

The probability of each factor of the product is assessed on the basis the relative fre-

quency of the values of the variables in the training set. The prior probability of event

duration T is calculated as the relative frequency of the duration in the training sample

(Total), as shown in (6.19).

P (T ) =F (T )

Total(6.19)

The conditional probability of the other factors are calculated from the joint probability,

which is estimated for each value of each variable as the joint relative frequency of the

values in the sample, as shown in (6.20-6.22).

P (Pf |T ) =F (Pf, T )

F (T )(6.20)

P (Sf |Pf, T ) =F (Sf, Pf, T )

F (Pf, T )(6.21)

P (Asp|Pf, Sf) =F (Asp, Pf, Sf)

F (Pf, Sf)(6.22)

In testing, the predicted time value is the one which is most likely, given the values of

the three aspect attributes observed in the test data:

duration class(instance) = arg maxt

P (t|pf, sf, asp) (6.23)

The conditional probability of each value t ∈ T is calculated applying the general con-

ditional probability rule, factorised according to the independence assumptions in the

Bayes’ net (see Figure 6.3), as shown in (6.24).

252


P (t|pf, sf, asp) =P (t) · P (pf |t) · P (sf |pf, t) · P (asp|pf, sf)∑t P (t) · P (pf |t) · P (sf |pf, t) · P (asp|pf, sf)

(6.24)

If one of the values of the three predictor variables is unseen in the training data, we

eliminate this variable from the evidence and calculate the conditional probability of t

given the remaining two variables, as shown in (6.25)-(6.27):

P (t|pf, sf) =

∑asp P (t) · P (pf |t) · P (sf |pf, t) · P (asp|pf, sf)∑

asp

∑t P (t) · P (pf |t) · P (sf |pf, t) · P (asp|pf, sf)

(6.25)

P (t|pf, asp) =

∑sf P (t) · P (pf |t) · P (Sf |pf, t) · P (asp|pf, sf)∑

sf


(6.26)

P (t|sf, asp) =

∑pf P (t) · P (pf |t) · P (sf |pf, t) · P (asp|pf, sf)∑

pf


(6.27)

In principle, the same variable elimination procedure can be applied also when two values

are unseen, but this was not necessary in our experiments.

6.4.2. Experimental evaluation

In the experimental evaluation, we train the model on a set of examples and then test

its predictions on an independent test set. In all the settings of the experiment, the

learning task is defined as supervised classification. The learning systems are trained on

a set of examples where the values of the target variable are observed together with the

values of the predictor variables. The predictions are then performed on a new set of

examples for which only the values of the predictor variables are observed.

The way the information from the predictors is used to predict values of the target

variable depends on the classification approach and on the machine learning algorithm

which is used. Some algorithms can be better suited for certain kinds of data than

others. (Hall et al. 2009). To determine which classification approach is the best

for our predictions, we perform several experiments. In addition to our Bayesian net

253


classifier, we test three more methods on two versions of the data. Our experimental

set-up is described in more detail in the following subsections.


We test our model on a set of examples with manually annotated event duration provided

by Pan et al. (2011) to which we assign verb aspect values acquired from Serbian verb

morphology through a parallel corpus as described in Section 6.3. The full set of data

which were used in the experiments is given in Appendix C.

Corpus and processing. The examples annotated with event duration are part of

the TimeBank corpus (Pustejovsky et al. 2003). The annotation of duration is, in

fact, added to the already existing TimeBank annotation. An example of an annotated

event is given in (6.28) (the mark-up language is XML). The part of the annotation in

bold face is the added duration information, the rest belongs to the original TimeBank

annotation.

(6.28) There’s nothing new on why the plane <EVENT eid=“e3”

class=“OCCURRENCE” lowerBoundDuration=“PT1S”

upperBoundDuration=“PT10S”>exploded</EVENT>.

In annotating event duration, the annotators were asked to assess a possible time span

over which the event expressed in a particular sentence can last. They were asked to

determine the lower and the upper bound of the span. We can see in the annotation of

the event in (6.28), for example, that the event of exploding is assessed to last between

one and ten seconds. Such annotations are provided for 2’132 event instances.

The agreement between three annotators is measured on a sample of around 10% of

instances. To measure the agreement, the seven time units (second, minute, hour, day,

week, month, year) were first converted into seconds. Then the mean value (between

the lower and upper bound) was taken as a single duration value. To account for the

different perception of the time variation in the short and long time spans, that is for

the fact that the difference between 3 and 5 seconds is perceived as more important

than the difference between 533 and 535 seconds, the values in seconds were converted

254


into their natural logarithms. The values on the logarithmic scale were then used to

define a threshold and divide all events into two classes: those that were assigned a

value less than 11.4 (which roughly corresponds to a day) were classified as short events

and the others were classified as long events. Pan et al. (2011) report a proportion of

agreement between the annotators of 0.877 on the two classes, which corresponds to the

κ-score of 0.75 when taking into account the expected agreement. This agreement can

be considered strong. It confirms that people do have common intuitions on how long

events generally last.

The events which are annotated by Pan et al. (2011) are expressed with different gram-

matical categories, including nouns (such as explosion), adjectives (such as worth), and

others. For testing our model, we select only those events which are expressed with

verbs, that is those instances which are assigned a verb tense value in an additional

layer of annotation, not shown in (6.28). We limit our data set to verbs because the the-

oretical notion of aspect, which we examine in our study, is essentially verbal. Category

changing is known to have considerable influence on certain elements of lexical structure

of words (Grimshaw 1990). Given the place it takes in the lexical representation of verbs

(Ramchand 2008), aspect can be expected to be one of the elements that are affected

by category changing. We thus use only the instances of events which are expressed by

verbs to avoid unnecessary noise in the data.

After eliminating non-verb events from the original Pan et al. (2011)’s corpus, the num-

ber of instances with annotated event duration were reduced to 1’121. We had to further

eliminate a number of these instance because we could only test our model on those verbs

for which we had acquired aspect information from the parallel corpus. Those are the

verbs which occur both in the TimeBank and in the MULTEXT-East corpus and which

are word-aligned with Serbian verbs (see Section 6.3). After eliminating the instances

for which we did not have Serbian alignments, we obtained the definitive set of data

which we used in the experiments, a total of 918 instances.

We follow Pan et al. (2011) in dividing all the events into two classes: short and long.

This decision is based on the fact that the inter-annotator agreement on more fine-

grained distinctions is much weaker than in the case of the binary classification. Pan

et al. (2011) also report agreement based on the overlap between the distributions of

255


duration values. This agreement score depends on the threshold defined for the overlap,

but reaches the kappa-score of 0.69 only when as little as 10% of the overlap is considered

agreement.

To transform the existing annotations into the two classes, we apply the procedure

which is described above: we convert all the time units into seconds, then transform

these values into natural logarithms and then set the threshold for dividing the events

into short and long to 11.4, which roughly corresponds to the length of one day.

Two versions of verb aspect annotation. Since human judgements on event dura-

tion agree much better for the binary classes than for more fine-grained distinctions, it

can be the case that verb aspect properties too are better represented as binary vari-

ables, instead of the ten-value representation shown in Section 6.3 (see Table 6.5). To

check whether a coarser representation of verb aspect is more useful for predicting the

two event duration classes, we perform experiments with two versions of the data. In the

first setting, we use the ten-value representation of verb aspect, as described in Section

6.3. In the second setting, we use only two values: high and low. We define the thresh-

old for dividing the values into these two classes as the median value of each variable

observed in the training data: the values are considered high if they are greater than 0.5

for Prefix, 0.3 for Suffix, and 0.7 for Aspect. Otherwise, the values are considered low.

Prefix Suffix Aspect Time0.8 0.5 0.8 short0.9 0.5 0.9 short0.8 0.8 0.2 short0.2 0.6 0.6 long0.6 0.0 0.6 long0.7 0.1 0.7 short0.8 0.2 0.3 short0.4 0.1 0.3 long0.4 0.4 0.4 long

Prefix Suffix Aspect Timehigh high high shorthigh high high shorthigh high low shortlow high low longhigh low low longhigh low high shorthigh low low shortlow low low longlow high low long

Table 6.6.: A sample of the two versions of data set with combined verb aspect and eventduration information: the version with ten-value predictor variables on theleft side, the version with two-value predictor variables on the right side.

As an illustration of the data which were used in the machine learning experiments, a

256


sample of instances is shown in Table 6.6. The first three columns of the two panels

contain the two versions of the representation of aspectual properties of English verbs

acquired from the morphological structure of their Serbian counterparts. The fourth

column contains human judgements of whether the events described by the verbs last

less or more than a day.

Comparison with other classifications methods. To assess whether our learning

approach is well-adapted to the kind of predictions that we make in our tests, we perform

the same classification of verbs into “short” and “long” but using different methods. In

all the methods with which we compare our Bayes net classifier, the representation of

aspect is the one described in Section 6.3.

The first classification method which we test is a simple rule-based classifier which does

not use the information about the morphological structure of Serbian verbs, but only

the binary classification into imperfective and perfective verbs. We classify as “short”

all English verbs which tend to be aligned with Serbian perfective verbs (these are the

verbs which are assigned 0.7 or more for the attribute Aspect as described in 6.3). The

other verbs are classified as “long”.

In addition to the Bayes net classifier and the simple rule-based classifier just described,

we train two machine learning algorithms: Decision Tree and Naıve Bayes classifiers.

Both algorithms are described in more detail in Section 3.4.1 in Chapter 3. We choose

these two algorithms because they are known to perform very well on a wide range of

learning tasks, despite their relative simplicity.

It is important to note that these two algorithms use our representation of aspect in

different ways. As discussed in Section 3.4.1, the Naıve Bayes approach is based on

the assumption that all the predictor variables are independent. Contrary to this, the

Bayesian net classifier includes the specific dependencies which we believe to exist in the

reality. This, if the dependencies are correctly identified, can be an advantage for the

Bayesian net classifier.

Another important difference is between Decision Tree on one side and both Bayesian

algorithms on the other. In the setting with the ten-value aspect attributes, values of

the variables are treated as real numbers in the Decision Tree experiments, while they

257


Mean accuracy score (%)Ten-value setting Two-value setting

Bayesian Net 83 79Decision Tree 83 79Naıve Bayes 79 76Binary aspect 70 70

Baseline 65 65

Table 6.7.: Results of machine learning experiments with 5-fold cross-validation (roughly80% of data for training and 20% for testing.

are treated as nominal values in both Bayesian experiments. To give an example, the

Decision Tree classifier “knows” that the value 0.8 is greater than 0.4, while Bayesian

classifiers treat these two numbers just as strings representing two different classes. This

difference can turn into an advantage for the Decision Tree classifier, because the true

nature of these values is ordinal.

We run the two classifiers using the packages available in the system Weka (Hall et al.

2009). The performance of all the three classifiers is reported in Table 6.7.

Training and test set, baseline. We evaluate the performance of the classifier on

the data set which consists of 918 instances for which both time annotation and cross-

linguistic data are available (see Section 6.4.2). The data set is split into the training

set (around 80%) and the test set (around 20%) using random choice of instances. To

account for the potential influence of the data split on the observed results, we perform

a five-fold cross-validation.3 In Table 6.7, we report the mean accuracy scores.

The baseline is defined as the proportion of the more common class in the data set.

Since 65% of instances belong to long events, if we assigned this label to all instances,

the accuracy of the classification would be 65%.

258



We can see in Table 6.7 that all four classifiers outperform the baseline in both settings.

The Bayesian Net and Decision Tree classifier perform better than Naıve Bayes classifier

(83% vs. 79%. mean accuracy score). This difference is statistically significant, as

determined by the t-test (t = 2.79, p < 0.01). The performance of all the three classifiers

is significantly better in the ten-value setting than in the two-value setting.

The differences in the performance of the four classifiers indicate that all the information

included in the model is useful. First, the simple rule-based classification gives the lowest

results, but still above the baseline. This indicates that there is a certain relationship

between the duration of an event and the perfective vs. imperfective interpretation of

Serbian verb which express it. However, the other methods, which combine the binary

division into perfective and imperfective with the morphological attributes, perform

much better.

The worst performing out of the three machine-learning classifiers is Naıve Bayes which

simplifies most the relationships between the properties, treating them as independent.

The fact that the Bayesian net classifier makes more correct predictions than Naıve

Bayes is likely to be due to the fact that the dependencies specified in the model express

true relationships between the structural elements. The fact that Decision Tree reaches

the same performance as our Bayesian net can be explained by the hierarchical nature

of the decisions taken by this classifier.

Finally, the consistent difference in the performance of all the three classifiers on the two

versions of data indicates that representing aspectual properties with only two values

is an oversimplification which results in more errors in predictions. Verb aspect clearly

cannot be described in terms of a single binary attribute, such as, for example, the

temporal boundedness, which is used in our study. Finer distinctions contain more

information which proves useful for predicting event duration.

3In five-fold cross validation, the data set is split in five portions. Each time the classifier is run oneportion is taken as the test set and all the other portions serve as the training test. The classifier isrun five time, once for each test set.

259



The results of the experiments can be interpreted in the light of several questions which

have been examined in our study. First of all, they clearly show that natural language

does encode those temporal properties of events about which people have common in-

tuitions and which are relevant for the representation of real world. The 83% accuracy

in predictions realised by the two best classifiers is much closer to the upper bound

of the performance on this task (the proportion of inter-annotator agreement of 0.877

measured by Pan et al. (2011) would correspond to an accuracy score of 87.7%) than

to the baseline (65%). This shows that the relevance of the linguistic encoding can be

learned and exploited by the systems for automatic identification of event duration.

Using Serbian language data to classify events expressed in English is based on the

assumption that verb aspect is an element of general structure of language and that

it has the same role in the languages where it is morphologically encoded as in the

languages which do not exhibit it in their observable morphological structure. This

assumption underlies all general linguistic accounts of verb aspect which we have taken

into consideration in our study. Successful transfer of the representation of aspect from

Serbian to English can be interpreted as a piece of empirical evidence in favour of this

assumption.

6.5.1. Aspectual classes

A careful analysis of the linguistic structure, guided by established theoretical notions,

proves useful in identifying the elements of the linguistic structure which are relevant

for temporal encoding. The temporal meaning of Serbian verb morphemes, which is

exploited by the systems in our experiments, becomes clear only in the context of the

general theory of verb aspect. The morphemes themselves are ambiguous and do not

constitute a temporal paradigm. However, the frequency distribution of the morphemes

in the cross-linguistic realisations of verbs clearly depends on the temporal properties

of the events described by the verbs and it does provide an observable indicator of the

temporal meaning of verbs.

260


The results of the experiments suggest that the three-attribute representation of aspect

which we propose, captures the relationships between the structural elements in Serbian

verbs which are relevant for time encoding. The simple binary classification is clearly

not an adequate level of aspectual classification. However, the set of ten values which

we have used, is not necessarily the best representation either. Our decision to group

the values into ten classes is arbitrary. It is possible that the classifiers do not use all

the ten values and that some of them are more informative than others. A systematic

approach to identifying the best representation of aspectual properties of verbs could help

identifying the elements of the lexical representation which interact with the temporal

boundedness in a systematic way to form fully specified aspectual classes. This would

improve our understanding of what kinds of aspectual classes exist and what kinds of

meaning they express.


Our results indicate that the kind of information elicited from human judges in anno-

tating event duration is represented mostly at the word level in the linguistic structure.

The event duration annotation which is used in our experiments is instance-based. The

annotators could assign different classes to different instances of the same verb form,

depending on the context of the form in a particular instance. This was not possible

in our approach to verb aspect. In order to transfer the acquired information from the

parallel corpus to the corpus which contains event annotation, we separated the aspect

representation from the context and tied it to verb types, assigning the same values to all

instances of the same type. The automatic classifiers were able to learn the relationship

between the two annotations despite this simplification, which indicates that the human

annotations were more influenced by the lexical properties of verbs than by the con-

text. This information could be useful in designing future approaches to identification

of temporal properties in natural language.

It should be noted that, even though we work with the representations at the word level,

our model is unlexicalised. We do not use the information contained in the idiosyncratic

lexical meaning of verbs, but more general elements of the representation shared by

different lexical items. As a consequence, a relatively small size of the training corpus

261


needed for learning both the empirical representation of verb aspect and its relation to

event duration. However, the word-level representation is, at the same time, a limitation

of our approach. Not all the information about event duration can be found at the

word level. A full model of linguistic encoding of time will have to take into account

observations at the higher levels of the structure of language too.

6.6. Related work

Pan et al. (2011) use the corpus described in Section 6.4.2 to perform machine learning

experiments. They define a set of features representing event instances which are used to

learn a classification of the events into short and long (see Section 6.4.2). The features

used in the event instance representation are: the lexical item which expresses the event,

its part-of-speech tag and lemma, the same information for the words in the immediate

context of the event item (in the window of size four), some syntactic dependencies of

the event item (such as its direct object, for example), and the WordNet (Fellbaum

1998) hypernyms of the event item. The classification is learned using three different

supervised algorithms: Decision Tree, Naıve Bayes, and Support Vector Machines. The

best performance is obtained by Support Vector Machines on the class of long events,

with an F-score of 0.82. Although the overall performance is not reported, it can be

expected to be lower than this score, given that the performance on the short events is

measured as an F-score of 0.65 (weighted average of the two scores is 0.75).

Our results are not directly comparable with the results of Pan et al. (2011). Our best

accuracy score corresponds to an overall F-score of 0.83 on both kinds of events, but

we do not use exactly the same training and test set. Since we have selected only the

instances of events expressed by verbs, we use only a portion of Pan et al. (2011)’s data

both for training and testing. Although we obtain a better score with a smaller data

set, we do not know what exactly causes the difference. A more thorough comparison

would be necessary to determine whether the task is easier on the instances which we

selected. This would justify our decision and underline the need for a different approach

to categories other than verbs. Otherwise, our approach could be judged as better, but

it should be extended to other categories.

262

6.6. Related work

Gusev et al. (2011) use predefined word patterns as indicators of event duration. One

of the patterns used, for example, is Past Tense + yesterday. If an event expressing

item shows a tendency to occur with this pattern, it can be taken to express a short

event in the sense of Pan et al. (2011). The occurrence data are extracted from the web

using the Yahoo! search engine. Gusev et al. (2011) train learning algorithms on the

instances where the event duration annotation is replaced by the pattern definitions. A

maximum entropy algorithm performs better than Support Vector Machines reaching

the best performance of 74.8%, which is not significantly different from the performance

on the hand annotated data set. Gusev et al. (2011) also try learning finer-grained

classes, but the accuracy scores are much lower (below 70%).

Feature analysis by both Pan et al. (2011) and Gusev et al. (2011) indicates that enrich-

ing the models with context information brings little or no improvement in the results,

which is in agreement with our own findings.

Williams and Katz (2012) explore other word patterns which indicate event duration for

classifying events into habitual and episodic. The data are extracted from a corpus of

Twitter messages and classified using a semi-supervised approach. The study suggests

that most verbs are used in both senses and proposes a lexicon of mean duration of

episodes and habits expressed by a set of verbs. These temporal quantifications, however,

are not directly evaluated against human judgments.

The work on verb aspect is mostly concerned with using elements of the context to

detect certain aspectual classes. The work of Siegel and McKeown (2000), for example,

addresses the aspectual classes proposed by Moens and Steedman (1988), showing, by

means of a regression analysis, that the context indicators which distinguish between

dynamic and stative events are different from the indicators which distinguish between

culminated and nonculminated events (the notion of a culminated event roughly cor-

responds to the notion of a temporally bounded event discussed in our study). Siegel

and McKeown (2000) also show that it is harder to distinguish between culminated and

non-culminated than between static and dynamic events. Kozareva and Hovy (2011)

propose a semi-supervised method for extracting word patterns correlated with a set

of aspectual classes, but their classes do not make reference to the classes discussed by

Siegel and McKeown (2000) nor to other classes argued for in the linguistic theory.

263


A possibility of cross-linguistic transfer of verb aspect through parallel corpora is ex-

plored by Stambolieva (2011), but the study is not conducted in the experimental frame-

work and does not report on automatic data processing.


In this study, we have explored the relationship between verb aspect, as an element

of the grammar of natural language, and the more general cognitive notion of event

duration. We have shown that this relationship can be explicitly formulated in terms

of a probabilistic model which predicts the duration of an event on the basis of the

aspectual representation of the verb which is used to express it. With the accuracy of

83%, the model’s predictions can be considered successful. The model’s accuracy score is

much closer to the upper bound, defined as the agreement between human classification

(88%), than to the baseline, defined as the proportion of the most frequent class (65%).

For the purpose of our experimental study, we have developed a quantitative represen-

tation of verb aspect which is based on the distribution of morphosyntactic realisations

of Serbian verbs in parallel English-Serbian instances of verbs. Contrary to other ap-

proaches to automatic identification of event duration, which have explored the observ-

able indicators at the syntactic and at the discourse level of linguistic structure, we have

identified observable indicators of event duration at the word level. We have shown that

a good proportion of temporal information which is implicitly understood in language

use is, in fact, contained in the grammar of lexical derivation of verbs in Serbian. This

information can be automatically acquired and ported across languages using parallel

corpora. The accuracy of the prediction based on our bi-lingual model is superior to the

best performing monolingual model.

264

7. Conclusion

In this dissertation, we have proposed a computational method for identifying gram-

matically relevant components of the meaning of verbs by observing the variation in

samples of verbs’ instances in parallel corpora. The core of our proposal is a formalisa-

tion of the relationship between the meaning of verbs and the variation, cross-linguistic

as well as language-internal, in their morphosyntactic realisations. We have used stan-

dard and Bayesian inferential statistics to provide empirical evidence of a number of

semantic components of the lexical representation of verbs which are grammatically rel-

evant because they play a role in the verbs’ predicate-argument structure. In particular,

we have shown that frequency distributions of morphosyntactic alternants in argument

alternations depend on the properties of events described by the alternating verbs.

Identifying grammatically relevant components of the meaning of verbs is one of the

core issues in linguistic theory due to the evident relationship between the meaning of

verbs and the structure of the clauses that they form. In order to understand how

basic clauses are structured, one needs to account for the differences in the number of

constituents which are realised. Such an account involves explaining why some clauses

are intransitive, some are transitive, and some are ditransitive. The explanation leads

to the lexical properties of the main verb which heads a clause. Intransitive clauses

are typically formed with the verbs such as go, swim, cough, laugh. Transitive clauses

are formed with verbs such as make, break, see. Ditransitive clauses are formed with

verbs such as give, tell. However, one needs to take into account also the fact that

one verb is rarely associated with only one type of a clause. It is much more often the

case that the same verb is associated with alternating clausal structures. For example,

the verb break can be realised in both a transitive clause (e.g. Adam broke the laptop)

and in a semantically related intransitive clause (e.g. The laptop broke). Alternative

265

7. Conclusion

morphosyntactic realisations of semantically equivalent units are not only found within

a single language, but also across languages. Although associating the meaning of verbs

with the types of clauses which they form is necessary for formulating the rules of clausal

structure, defining precise rules to link the elements of a phrase structure to the elements

of the lexical representation of verbs proves to be a challenging task.

The work on the interface between the lexicon and the rules of grammar has resulted

in numerous proposals regarding the grammatically relevant elements of the lexical rep-

resentation of verbs. It is widely accepted that the meaning of a verb is related to the

grammar of a clause through a layer in the lexical representation of verbs which is usu-

ally called the predicate-arguments structure. However, the views of what exactly the

elements of the predicate-argument structure are differ very much.

The nature of the predicate-argument relations in the representation of the meaning

of verbs has been described in various frameworks, starting with the naıve analyses of

semantic roles of verbs’ arguments (Fillmore 1968; Baker et al. 1998) to more general

approaches based on semantic decomposition of the predicate-argument relations. Sev-

eral attributes of the meaning of verbs have been proposed as relevant for the predicate

arguments structure. It has been argued, for example, that these attributes include vo-

lition (Dowty 1991) or, more generally, mental state (Reinhart 2002), change (Dowty

1991; Reinhart 2002; Levin and Rappaport Hovav 1994), causation (Talmy 2000;

Levin and Rappaport Hovav 1994). The fact that the morphosyntactic realisations of

verbs are influenced by the values of these attributes makes these semantic components

grammatically relevant (Pesetsky 1995; Levin and Rappaport Hovav 2005). For exam-

ple, a verb can be expected to form intransitive clauses if it describes an event which

does not involve somebody’s volition. The attributes can interact between themselves

and also with other factors, which results in a complex relationship between the lexical

representation of verbs and the structure of a clause.

In more recent accounts, the components of the predicate-argument structure have been

reinterpreted in terms of temporal decomposition of events described by verbs (Krifka

1998; Ramchand 2008; Arsenijevic 2006). The notion of causation, for example, is iden-

tified with the notion of temporal precedence, while the notion of change is reanalysed

as a result. The structural elements proposed in the temporal account of the predicate-

266

7.1. Theoretical contribution

argument relations are usually called sub-events. The defined relations hold between the

sub-events.

We have proposed an empirical statistical method to test theoretical proposals regarding

the relationship between the lexical structure of verbs and the structure of a clause on

a large scale. Following the influential study on large-scale semantic classification of

verbs (Levin 1993), we have based our approach on the assumption that the meaning

of a verb determines the syntactic variation in the structure of the clauses that it forms

and that, therefore, the grammatically relevant components of the meaning of a verb

can be identified by observing the variation in its syntactic behaviour. The validity

of this general assumption has already been tested in the context of automatic verb

classification (Merlo and Stevenson 2001). In this dissertation, we have formulated

and tested experimentally a number of specific hypotheses showing that the frequency

distribution of syntactic alternants in the morphosyntactic realisations of verbs can be

predicted from some particular properties of the meaning of verbs. We have applied our

approach to two general properties of events which have been widely discussed in the

recent literature: causation and temporal structure. The contribution of our work with

respect to the existing work is both theoretical and methodological.

7.1. Theoretical contribution

With respect to previous theoretical approaches to the predicate-argument structure

of verbs, the main novelty of our work is the demonstrated grammatical relevance of

certain attributes of the meaning of verbs. Using statistical inference, we have formalised

the relationship between the meaning of verbs, their use, represented by the frequency

distribution of their instances in a corpus, and their formal properties (such as causative

or aspectual marking), showing how the three sources of data can be combined in a

unified account of the interface between the lexicon and the grammar.

In an analysis of the relationship between the kind of causation and the variation in mor-

phosyntactic realisations of light verb constructions, we have found empirical evidence of

the presence of two force-dynamics schemata in light verbs. The meaning of light verbs

such as take can be described as self-oriented (Talmy 2000; Brugman 2001) because the

267

7. Conclusion

dynamics of the event is oriented towards its causer (or agent). As opposed to this, the

meaning of light verbs such as make can be described as directed because the dynamics

of the event is not oriented towards the causer, but towards another participant in the

event. Our experiments have shown that the frequency distribution of cross-linguistic

morphosyntactic alternants of light verbs depend on their force-dynamics schemata.

In an analysis of cross-linguistic morphosyntactic realisations of lexical causatives, we

have taken a closer look into the notion of external causation (Haspelmath 1993; Levin

and Rappaport Hovav 1994; Alexiadou 2010). We have argued, based on the results of

a series of experiments, that the likelihood of an external causer in an event described

by a verb is a semantic property which underlies two correlated frequency distributions:

the distribution of morphological marking on lexical causatives across a wide range of

languages and the distribution of clause types in a sample of verb instances in any single

language. The contribution of this piece of work is twofold. First, we have shown that

there is a relationship between a semantic attribute of lexical causatives and their mor-

phosyntactic form. Specifically, the observed variation in the cross-linguistic realisations

of a verb depends on the likelihood of external causation of the event described by the

verb. Second, we have shown that the likelihood of external causation can be estimated

for a wide range of verbs by means of a statistical model.

The temporal structure of events is analysed in the third case study. The main contri-

bution of this study is the established relationship between formal aspectual marking on

a verb and the duration of the event described by the verb. More specifically, we have

designed a statistical model which predicts the duration of an event described by an

English verb on the basis of the observed frequency distribution of formal morphosyn-

tactic aspectual markers in the aligned Serbian verbs. In an experimental evaluation,

the model is shown to make better predictions than the best performing monolingual

model.

We have developed corpus-based measures of the values of the three semantic attributes

of verbs which we have studied. These values are calculated automatically and they can

be assigned to a large number of verbs.

268

7.2. Methodological contribution

7.2. Methodological contribution

The main methodological contribution of this dissertation is that it combines theoretical

linguistic goals with the sound modelling and experimental methodology developed in

computational linguistics. The methodology which we have used in this dissertation is

not new in itself, but its application to testing theoretical hypotheses is novel in three

ways.

First, while frequency distributions of syntactic realisations of verbs have been exten-

sively studied and used in the context of developing practical applications in the domain

of automatic natural language processing, this kind of evidence is not commonly used

in theoretical linguistics. In our experiments, we have demonstrated that a statistical

analysis of a large number of verb instances can be used to study structural components

in the lexical representation of verbs. We have quantified and measured the semantic

phenomena which we have studied using the methods and the techniques developed in

natural language processing. We have used statistical models and tests to capture the

generalisations in large data sets. We have estimated the parameters of the models by

applying machine learning techniques. We have tested and explicitly evaluated the pre-

dictions of the models. These methods constitute the standard experimental paradigm

in computational linguistics. In this dissertation, we have shown that their application

to addressing theoretical questions can lead to extending our knowledge about language.

Combining various sources of data in a large-scale analysis can shed some new light on

the nature of the interface between the lexicon and the grammar, which involves complex

interactions of multiple factors.

Second, the data sets which are used in the standard linguistic approaches are usually

much smaller than those which are used in our experiments. The methodological ad-

vantage of large data sets is that they are more likely to be representative of linguistic

phenomena than small samples which are manually analysed. By using computer-based

language technology, we can now observe the variation in the use of linguistic units

on a large scale, applying inductive reasoning in defining generalisations. In this dis-

sertation, we have shown how the tools and resources developed in natural language

processing can be used to compose large experimental data sets for theoretical linguistic

research. We have used existing parallel corpora, an automatic alignment tool, syntactic

269

7. Conclusion

parsers, morphological analysers, as well as our own scripts for automatic extraction of

the experimental data from parallel corpora. With the rapidly developing language tech-

nology, such resources can be expected to grow and to be increasingly available in the

future. The data and the tools accumulated in developing language technology represent

extremely rich new resource for future linguistic research.

Third, we have extended the corpus-based quantitative approach to linguistic analysis to

the cross-linguistic domain. This is a necessary step for formulating generalisations which

hold across languages. We have achieved this by collecting data from parallel corpora.

We have shown that parallel corpora represent a valuable source of information for a

systematic study of the structural sources of cross-linguistic microvariation, despite the

fact that the observed variation can be influenced by some non-linguistic factors (such

as translators’ choice) as well.

7.3. Directions for future work

We define the directions for continuing the work presented in this dissertation in two

ways. On the one hand, our approach can be extended to include more languages and

to more complex modelling. On the other hand, our findings have opened new questions

which could be pursued further in future work.

Although our approach is cross-linguistic in the sense that we analyse the data from at

least two languages in all our experiments, our data come from only a few languages: En-

glish, German, and Serbian. We have used only a small sample of languages because the

focus of our work has been on developing and testing the methodology of cross-linguistic

corpus-based linguistic research. Applying the methods proposed in this dissertation to

a larger sample of languages is a natural next step in future research.

Increasing the number of languages included in an analysis would enrich the data sets not

only because more instances of linguistic phenomena would be analysed, but also because

more linguistic information can be automatically extracted. For example, morphological

marking, which is often not available in English, can be extracted from other languages.

Although we have not used morphological marking in a systematic way, the results

270

7.3. Directions for future work

of our experiments suggest that it can be a valuable source of information to study

various elements of the grammar of language, which is in accordance with some recent

broad typological studies (Bickel et al. 2013; To appear). Parallel corpora of numerous

languages already exist (for example, the current version of the corpus Europarl contains

21 languages) and they are constantly growing.

Since statistical modelling has not been widely used in theoretical linguistic research so

far, the focus of this dissertation has been on demonstrating how statistical inference

can be used to address theoretical issues. To this end, we have formulated relatively

narrow theoretical questions which could have been addressed using simple statistical

and computational approaches. This allowed us to establish a straightforward relation-

ship between theoretical notions which we studied and the components of the models.

However, our approach can be extended to more general questions involving more fac-

tors. This can be done by applying more advanced modelling approaches such as those

which are currently being proposed in computational linguistics and in other disciplines

dealing with large-scale data analysis.

By analysing the data in our experiments, we have noticed several phenomena which

call for further investigation, but which we could not address directly because this work

would fall out of the scope of the dissertation. Such a phenomenon is, for example,

the fact that nouns are aligned better than verbs in automatic alignment in general.

It would be worth exploring in future work whether this fact can be related to some

known distributional differences between these two classes or not. It might also mean

that nominal lexical items are more stable across languages than verbal ones.

Another phenomenon which would be worth exploring in future research became evident

while we were studying the data on lexical causatives. We have noticed that the quantity

of anticausative morphological marking varies in European languages. The number of

lexical causatives which are attached a reflexive particle in the citation form, such as

sich offnen ’open’ in German, varies across European languages. There are, for example,

many more such verbs in Serbian than in German, while there are almost none in English.

Possible explanation for this variation is the difference in the morphological richness

between the three languages, given that Serbian is usually considered morphologically

richer than German, and German richer than English. Based on our results, this marking

271

7. Conclusion

could be expected to be related to the likelihood of external causation too. The verbs

which describe an event with a low likelihood of an external causer are expected to occur

without a marker more than the verbs describing an event with a high likelihood of an

external causer. The morphological markings should, thus, be distributed in a continuous

fashion over the scale of likelihood of external causation, covering different portions of

the scale in different languages. Addressing these relations directly in an experiment

might result in new findings pointing to some structural constraints on cross-linguistic

variation.

Finally, in the study of temporal properties of the meaning of verbs, we have proposed

a quantitative representation of verb aspect classes based on frequency distribution of

morphological marking in Serbian verbs. This representation proved useful for the goals

of our experiments. However, we have not fully examined the theoretical aspects of our

proposal. What remains as an open question for future research is the source of the

quantities which have been observed in our experiments as the values of the aspectual

attributes. Exploring this question further could point to new findings on how mor-

phological marking patterns can be used for determining which aspectual classes exist

in language and what their meaning is. This should provide a clearer picture of the

semantic representation of time in language in general.

272

Bibliography

Steven Abney. Data-intensive experimental linguistics. Linguistic Issues in Language

Technology — LiLT, 6(2):1–30, 2011.

Enoch Olade Aboh. Clause structure and verb series. Linguistic Inquiry, 40(1):1–33,

2009.

Artemis Alexiadou. On (anti-)causative alternations, 2006a. Presentation, Ecole

d’automne de linguistique, Paris.

Artemis Alexiadou. On the morpho-syntax of (anti-)causative verbs. In Malka Rappa-

port Hovav, Edit Doron, and Ivy Sichel, editors, Syntax, Lexical Semantics and Event

Structure, pages 177–203, Oxford, 2010. Oxford University Press.

Artemis Alexiadou, Elena Anagnostopoulou, and Florian Schafer. The properties of

anticausatives crosslinguistically. In Mara Frascarelli, editor, Phases of Interpretation,

pages 187–212, Berlin, New York, 2006. Mouton de Gruyter.

Alex Alsina. A theory of complex predicates: evidence from causatives in Bantu and

Romance. In Alex Alsina, Joan Bresnan, and Peter Sells, editors, Complex predicates,

pages 203–247, Stanford, California, 1997. CSLI Publications.

Boban Arsenijevic. Inner aspect and telicity: The decompositional and the quantifica-

tional nature of eventualities at the syntax-semantics interface. LOT, Utrecht, 2006.

Boban Arsenijevic. Slavic verb prefixes are resultative. Cahiers Chronos, 17:197–213,

2007.

Harald Baayen. Analyzing Linguistic Data. A Practical Introduction to Statistics using

R. Cambridge University Press, Cambridge, 2008.

273

Bibliography

Collin F. Baker, Charles J. Fillmore, and John B. Lowe. The berkeley framenet project.

In Proceedings of the 36th Annual Meeting of the Association for Computational Lin-

guistics and 17th International Conference on Computational Linguistics, pages 86–90,

Montreal, Canada, 1998. ACL / Morgan Kaufmann Publishers.

Mark Baker. Thematic roles and syntactic structure. In Liliane Haegeman, editor,

Elements of Grammar, pages 73–137, Dordrecht, 1997. Kluwer.

Mark C. Baker. Incorporation — A Theory of Grammatical Function Changing. The

University of Chicago Press, Chicago, London, 1988.

Mark C. Baker. The atoms of language. Basic Books, New York, 2001.

Marco Baroni and Silvia Bernardini. A new approach to the study of translationese:

Machine-learning the difference between original and translated text. Literary and

Linguistic Computing, 21(3):259–274, 2006.

Marco Baroni and Alessandro Lenci. Distributional memory: A general framework for

corpus-based semantics. Computational Linguistics, 36(4):673–722, 2010.

John Beavers. Argument/Oblique Alternations and the Structure of Lexical Meaning.

PhD thesis, Stanford University, 2006.

Douglas Biber. Variation across Speech and Writing. Cambridge University Press,

Cambridge, 1991.

Theresa Biberauer, editor. The Limits of Syntactic Variation, Amsterdam, 2008. John

Benjamins.

Balthasar Bickel, Giorgio Iemmolo, Taras Zakharko, and Alena Witzlack-Makarevich.

Patterns of alignment in verb agreement. In Dik Bakker and Martin Haspelmath,

editors, Languages across boundaries: studies in the memory of Anna Siewierska,

pages 15–36. De Gruyter Mouton, Berlin, 2013.

Balthasar Bickel, Taras Zakharko, Lennart Bierkandt, and Alena Witzlack-Makarevich.

Semantic role clustering: an empirical assessments of semantic role types in non-

default case assignment, To appear.

274

Bibliography

Claire. Bonial, William. Corvey, Martha. Palmer, Volha V. Petukhova, and Harry Bunt.

A hierarchical unification of lirics and verbnet semantic roles. In Semantic Computing

(ICSC), 2011 Fifth IEEE International Conference on, pages 483–489, Sept 2011. doi:

10.1109/ICSC.2011.57.

Gerlof Bouma, Lilja Øvrelid, and Jonas Kuhn. Towards a large parallel corpus of cleft

constructions. In Proceedings of the Seventh conference on International Language

Resources and Evaluation (LREC’10), Valletta, Malta, 2010. European Language Re-

sources Association.

Melissa Bowerman and William Croft. The acquisition of the English causative alterna-

tion. In Melissa Bowerman and Penelope Brown, editors, Crosslinguistic perspectives

on argument structure: Implications for learnability, pages 279–306, New York, NY,

2008. Lawrence Erlbaum Associates.

Michael R. Brent. From grammar to lexicon: Unsupervised learning of lexical syntax.

Computational Linguistics, 19(3):243–262, 1993.

Joan Bresnan. Is syntactic knowledge probabilistic? Experiments with the English dative

alternation. In Sam Featherston and Wolfgang Sternefeld, editors, Roots: Linguistics

in Search of Its Evidential Base, Studies in Generative Grammar, pages 77–96, Berlin,

2007. Mouton de Gruyter.

Joan Bresnan and Tatiana Nikitina. The gradience of the dative alternation. In Linda

Uyechi and Lian Hee Wee, editors, Reality Exploration and Discovery: Pattern Inter-

action in Language and Life, pages 161–184, Stanford, 2009. CSLI Publications.

Ted Briscoe and John Carroll. Automatic extraction of subcategorization from corpora.

In Proceedings of the 5th ACL Conference on Applied Natural Language Processing,

pages 356–363, 1997.

Peter F. Brown, Stephen A. Della-Pietra, Vincent J. Della-Pietra, and Robert L. Mercer.

The mathematics of statistical machine translation. Computational Linguistics, 19(2):

263–313, 1993.

Claudia Brugman. Light verbs and polysemy. Language Science, 23:551–578, 2001.

275

Bibliography

Aljoscha Burchardt, Katrin Erk, Anette Frank, Andrea Kowalski, Sebastian Pado, and

Manfred Pinkal. Using frameNet for the semantic analysis of German: Annotation,

representation and automation. In Hans Boas, editor, Multilingual FrameNets in

computational lexicography, pages 209–244. Mouton de Guyter, 2009.

Lou Burnard. Reference guide for the British National Corpus (XML edition), 2007.

URL http://www.natcorp.ox.ac.uk/XMLedition/URG/.

Miriam Butt and Wilhelm Geuder. On the (semi)lexical status of light verbs. In Norbert

Corver and Henk van Riemsdijk, editors, Semilexical Categories: On the content of

function words and the function of content words, pages 323–370, Berlin, 2001. Mouton

de Gruyter.

Xavier Carreras and Lluis Marquez. Introduction to the CoNLL-2005 shared task: Se-

mantic role labeling. In Proceedings of the 9th conference on computational natural

language learning (CONLL), pages 152–164, Ann Arbor, 2005. Association for com-

putational linguistics.

Noam Chomsky. Remarks on nominalization. In Roderick Jacobs and Peter Rosen-

baum, editors, Readings in English Transformational Grammar, Waltham, MA, 1970.

Blaisdell.

Noam Chomsky. Knowledge of language: its nature, origin and use. Praeger, New York,

1986.

Noam Chomsky. The minimalist program. MIT Press, Cambridge, Massachusetts, 1995.

Kenneth W. Church and Patrick Hanks. Word association norms, mutual information,

and lexicography. Computational Linguistics, 16(1):22–29, 1990.

Trevor Cohn and Mirella Lapata. Machine translation by triangulation: Making effec-

tive use of multi-parallel corpora. In Proceedings of the 45th Annual Meeting of the

Association of Computational Linguistics, pages 728–735, Prague, Czech Republic,

June 2007. Association for Computational Linguistics.

Chris Collins. Argument sharing in serial verb constructions. Linguistic Inquiry, 28:

461–497, 1997.

276

Bibliography

Michael Collins, Philipp Koehn, and Ivona Kucerova. Clause restructuring for statistical

machine translation. In Proceedings of the Annual Meeting of the Association for

Computational Linguistics (ACL), pages 531–540, Ann Arbor, 2005. Association for

Computational Linguistics.

Michael Cysouw and Bernhard Walchli, editors. Parallel Texts. Using Translational

Equivalents in Linguistic Typology, volume Theme issue in Sprachtypologie und Uni-

versalienforschung (STUF) 60(2), 2007. Akademie Verlag GMBH.

David Dowty. Thematic proto-roles and argument selection. Language, 67(3):547–619,

1991.

David R Dowty. Word meaning and Montague grammar: the semantics of verbs and

times in generative semantics and in Montague’s PTQ. D. Reidel, cop., Dordrecht,

Boston, 1979.

David R. Dowty. The effects of aspectual class on the temporal structure of discourse:

semantics or pragmatics. Linguistics and Philosophy, 9:37–61, 1986.

Tomaz Erjavec. MULTEXT-East version 4: Multilingual morphosyntactic specifications,

lexicons and corpora. In Nicoletta Calzolari (Conference Chair), Khalid Choukri,

Bente Maegaard, Joseph Mariani, Jan Odijk, Stelios Piperidis, Mike Rosner, and

Daniel Tapias, editors, Proceedings of the Seventh conference on International Lan-

guage Resources and Evaluation (LREC’10), pages 2544–2547, Valletta, Malta, 2010.

European Language Resources Association (ELRA).

Afsaneh Fazly. Automatic acquisition of lexical knowledge about multiword predicates.

PhD thesis, University of Toronto, 2007.

Christiane Fellbaum, editor. WordNet: An Electronic Lexical Database. MIT Press,

Cambridge, Mass., 1998.

Charles Fillmore. The case for case. In Emmon Bach and Robert T. Harms, editors,

Universals in linguistic theory, pages 1–88, New York, 1968. Holt, Rinehart and Win-

ston.

277

Bibliography

Charles J. Fillmore. Frame semantics. In Linguistics in the Morning Calm, pages 111–

137, Seoul, 1982. Hanshin Publishing Co.

Pascale Fung, Zhaojun Wu, Yongsheng Yang, and Dekai Wu. Learning bilingual semantic

frames: Shallow semantic parsing vs. semantic role projection. In 11th Conference on

Theoretical and Methodological Issues in Machine Translation (TMI 2007), pages 75–

84, Skovde, Sweden, 2007.

Nikhil Garg and James Henderson. Unsupervised semantic role induction with global

role ordering. In Proceedings of the 50th Annual Meeting of the Association for Com-

putational Linguistics (Volume 2: Short Papers), pages 145–149, Jeju Island, Korea,

July 2012. Association for Computational Linguistics. URL http://www.aclweb.

org/anthology/P12-2029.

Matthew Gerber and Joyce Y. Chai. Semantic role labeling of implicit arguments for

nominal predicates. Computational Linguistics, 38(4):755–798, 2012.

Daniel Gildea and Daniel Jurafsky. Automatic labeling of semantic roles. Computational

Linguistics, 28(3):245–288, 2002.

Gregory Grefenstette and Simone Teufel. Corpus-based method for automatic identifi-

cation of support verbs for nominalization. In Proceedings of the 7th Meeting of the

European Chapter of the Association for Computational Linguistics, pages 98–103,

Dublin, Irland, 1995. Association for Computational Linguistics.

H. Paul Grice. Logic and conversation. In Peter Cole and Jerry L. Morgan, editors,

Syntax and Semantics 3: Speech Acts, pages 41–58, New York, 1975. Academic Press.

Jane Grimshaw. Argument Structure. MIT Press, Cambridge, Mass., 1990.

Jane Grimshaw and Armin Mester. Light verbs and theta-marking. Linguistic Inquiry,

19:205–232, 1988.

Andrey Gusev, Nathaniel Chambers, Pranav Khaitan, Divye Khilnani, Steven Bethard,

and Dan Jurafsky. Using query patterns to learn the durations of events. In IEEE

IWCS-2011, 9th International Conference on Web Service, pages 145–155, Oxford,

UK, 2011. Institute of Electrical and Electronics Engineers (IEEE ).

278

Bibliography

Jan Hajic, Massimiliano Ciaramita, Richard Johansson, Daisuke Kawahara,

Maria Antonia Martı, Lluıs Marquez, Adam Meyers, Joakim Nivre, Sebastian Pado,

Jan Stepanek, Pavel Stranak, Mihai Surdeanu, Nianwen Xue, and Yi Zhang. The

CoNLL-2009 shared task: Syntactic and semantic dependencies in multiple languages.

In Proceedings of the Thirteenth Conference on Computational Natural Language

Learning (CoNLL 2009): Shared Task, pages 1–18, Boulder, Colorado, June 2009.

Association for Computational Linguistics.

Kenneth Hale and Samuel Jay Keyser. On argument structure and the lexical represen-

tation of syntactic relations. In Kenneth Hale and Samuel Jay Keyser, editors, The

View from Building 20, pages 53–110. MIT Press, 1993.

Mark Hall, Eibe Frank, Geoffrey Holmes, Bernhard Pfahringer, Peter Reutemann, and

Ian H. Witten. The WEKA data mining software: An update. SIGKDD Explorations,

11(1), 2009.

Martin Haspelmath. More on typology of inchoative/causative verb alternations. In

Bernard Comrie and Maria Polinsky, editors, Causatives and transitivity, volume 23,

pages 87–121, Amsterdam/Philadelphia, 1993. John Benjamins Publishing Co.

John A. Hawkins. A comparative typology of English and German : unifying the con-

trasts. Croom Helm, London ; Sydney, 1986.

James Henderson, Paola Merlo, Gabriele Musillo, and Ivan Titov. A latent variable

model of synchronous parsing for syntactic and semantic dependencies. In Alex Clark

and Kristina Toutanova, editors, Proceedings of the Twelfth Conference on Compu-

tational Natural Language Learning (CONLL 2008), page 178–182, Manchester, UK,

2008.

Rebecca Hwa, Philip Resnik, Amy Weinberg, and Okan Kolak. Evaluation translational

correspondance using annotation projection. In Proceedings of the 40th Annual Meet-

ing of the Association for Computational Linguistics, pages 392–399, Philadelphia,

PA, 2002. Association for Computational Linguistics.

Jena D. Hwang, Archna Bhatia, Claire Bonial, Aous Mansouri, Ashwini Vaidya, Ni-

anwen Xue, and Martha Palmer. PropBank annotation of multilingual light verb

279

Bibliography

constructions. In Proceedings of the Fourth Linguistic Annotation Workshop, pages

82–90, Uppsala, Sweden, July 2010. Association for Computational Linguistics.

Ray Jackendoff. X syntax : a study of phrase structure. MIT Press, Cambridge Mass.,

1977.

Eric Joanis and Suzanne Stevenson. A general feature space for automatic verb clas-

sification. In Proceedings of The 10th Conference of the European Chapter of the

Association for Computational Linguistics (EACL 2003), pages 163–170, Budapest,

Hungary, 2003. Association for Computational Linguistics.

Eric Joanis, Suzanne Stevenson, and David James. A general feature space for automatic

verb classification. Natural Language Engineering, 14(3):337–367, 2008.

Richard Kayne. The Oxford Handbook of Comparative Syntax, chapter Some notes on

comparative syntax, with special reference to English and French. Oxford University

Press, 2005.

Kate Kearns. Light verbs in english. Manuscript, 2002.

Karin Kipper Schuler. VerbNet: A broad-coverage, comprehensive verb lexicon. PhD

thesis, University of Pennsylvania, 2005.

Dan Klein. The unsupervised learning of natural language structure. PhD thesis, Stanford

University, 2005.

Philipp Koehn. Europarl: A parallel corpus for statistical machine translation. In

Proceedings of MT Summit 2005, Phuket, Thailand, 2005.

Zornitsa Kozareva and Eduard Hovy. Learning temporal information for states and

events. In Proceedings of the Workshop on Semantic Annotation for Computational

Linguistic Resources (ICSC 2011), Stanford, 2011.

Manfred Krifka. The origins of telicity. In Susan Rothstein, editor, Events and Grammar,

pages 197–235, Dordrecht, 1998. Kluwer.

280

Bibliography

Cvetana Krstev, Dusko Vitas, and Tomaz Erjavec. MULTEXT-East resources for Ser-

bian. In Proceedings of 8th Informational Society - Language Technologies Conference,

IS-LTC, pages 108–114, Ljubljana, Slovenia, 2004.

Jonas Kuhn. Experiments in parallel-text based grammar induction. In Proceedings of

the 42nd Meeting of the Association for Computational Linguistics (ACL’04), Main

Volume, pages 470–477, Barcelona, Spain, July 2004.

Joel Lang and Mirella Lapata. Unsupervised semantic role induction via split-merge

clustering. In Proceedings of the 49th Annual Meeting of the Association for Com-

putational Linguistics: Human Language Technologies, pages 1117–1126, Portland,

Oregon, USA, June 2011. Association for Computational Linguistics. URL http:

//www.aclweb.org/anthology/P11-1112.

Maria Lapata. Acquiring lexical generalizations from corpora: A case study for diathe-

sis alternations. In Proceedings of the 37th Annual Meeting of the Association for

Computational Linguistics, pages 397–404, College Park, Maryland, USA, June 1999.


Mirella Lapata and Chris Brew. Verb class disambiguation using informative priors.

Computational Linguistics, 30(1):45–73, 2004.

Richard K. Larson. On the double object construction. Linguistic Inquiry, 19:335–391,

1988.

Alex Lascarides and Nicholas Asher. Temporal interpretation, discourse relations and

commonsense entailment. Linguistics and Philosophy, 16(5):437–493, 1993.

Beth Levin. English verb classes and alternations : a preliminary investigation. The

University of Chicago Press, Chicago, 1993.

Beth Levin and Malka Rappaport Hovav. A preliminary analysis of causative verbs in

English. Lingua, 92:35–77, 1994.

Beth Levin and Malka Rappaport Hovav. Unaccusativity : at the syntax-lexical semantics

interface. MIT Press, Cambridge, Mass., 1995.

281

Bibliography

Beth Levin and Malka Rappaport Hovav. Argument realization. Cambridge University

Press, Cambridge, 2005.

Stephen C. Levinson. Pragmatics. Cambridge Textbooks in Linguistics. Cambridge

University Press, Cambridge, 1983.

Jianguo Li and Chris Brew. Which are the best features for automatic verb classification.

In Proceedings of ACL-08: HLT, pages 434–442, Columbus, Ohio, June 2008. Associ-

ation for Computational Linguistics. URL http://www.aclweb.org/anthology/P/

P08/P08-1050.

Edward Loper, Szu-Ting Yi, and Martha Palmer. Combining lexical resources: Mapping

between propbank and verbnet. In Proceedings of the 7th International Workshop on

Computational Linguistics, Tilburg, the Netherlands, 2007.

Catherine Macleod, Ralph Grishman, Adam Meyers, Leslie Barrett, and Ruth Reeves.

NOMLEX: A lexicon of nominalizations. In In Proceedings of Euralex98, pages 187–

193, 1998.

Christopher D. Manning. Automatic acquisition of a large subcategorization dictionary

from corpora. In Proceedings of the 31st Annual Meeting of the Association for Com-

putational Linguistics, pages 235–242, Columbus, Ohio, USA, June 1993. Association

for Computational Linguistics.

Mitchell Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. Building a large

annotated corpus of english: the penn treebank. Computational Linguistics, 19(2):

313–330, 1994.

Rafael Marın and Louise McNally. Inchoativity, change of state, and telicity: Evidence

from Spanish reflexive psychological verbs. Natural Language and Linguistic Theory,

29:467–502, 2011.

Lluıs Marquez, Xavier Carreras, Kenneth C. Litkowski, and Suzanne Stevenson. Seman-

tic role labeling: An introduction to the special issue. Computational Linguistics, 34

(2):145–159, 2008.

282

Bibliography

Diana McCarthy and Anna Korhonen. Detecting verbal participation in diathesis al-

ternations. In Proceedings of the 36th Annual Meeting of the Association for Compu-

tational Linguistics and 17th International Conference on Computational Linguistics,

Volume 2, pages 1493–1495, Montreal, Quebec, Canada, August 1998. Association for


Gail McKoon and Talke Macfarland. Externally and internally caused change of state

verbs. Language, 76(4):833–858, 2000.

Paola Merlo and Gabriele Musillo. Semantic parsing for high-precision semantic role

labelling. In Proceedings of the 12th conference on computational natural language

learning (CONLL), pages 1–8, Manchester, 2008. Association for Computational Lin-

guistics.

Paola Merlo and Susanne Stevenson. Automatic verb classification based on statistical

distribution of argument structure. Computational Linguistics, 27(3):373–408, 2001.

Paola Merlo and Lonneke van der Plas. Abstraction and generalization in semantic role

labels: PropBank, VerbNet or both? In Proceedings of the Joint Conference of the 47th

Annual Meeting of the ACL and the 4th International Joint Conference on Natural

Language Processing of the AFNLP, pages 288–296, Singapore, 2009. Association for


Paola Merlo, Suzanne Stevenson, Vivian Tsang, and Gianluca Allaria. A multilin-

gual paradigm for automatic verb classification. In Proceedings of 40th Annual

Meeting of the Association for Computational Linguistics, pages 207–214, Philadel-

phia, Pennsylvania, USA, July 2002. Association for Computational Linguistics. doi:

10.3115/1073083.1073119. URL http://www.aclweb.org/anthology/P02-1027.

Natasa Milicevic. The lexical and superlexical verbal prefix iz- and its role in the stacking

of prefixes. Nordlyd, 32(2):279–300, 2004.

Tom T. Mitchell. Machine Learning. McGraw-Hill, Boston, Mass., 1997.

Marc Moens and Mark Steedman. Temporal ontology and temporal reference. Compu-

tational Linguistics, 14(2):15–28, June 1988.

283

Bibliography

Jacques Moeschler and Anne Reboul. Dictionnaire encyclopedique de pragmatique. Ed.

du Seuil, Paris, 1994.

Paola Monachesi, Gerwert Stevens, and Jantine Trapman. Adding semantic role an-

notation to a corpus of written Dutch. In Proceedings of the Linguistic Annotation

Workshop (LAW), pages 77–84, Prague, Czech Republic, 2007. Association for Com-

putational Linguistic.

Joakim Nivre, Johan Hall, Jens Nilsson, Chanev Atanas, Gulesen Eryigit, Sandra

Kubler, Svetoslav Marinov, and Erwin Marsi. MaltParser: A language-independent

system for data-driven dependency parsing. Natural Language Engineering, 13(2):

95–135, 2007.

Franz Josef Och and Hermann Ney. A systematic comparison of various statistical

alignment models. Computational Linguistics, 29(1):19–52, 2003.

Sebastian Pado. Cross-Lingual Annotation Projection Models for Role-Semantic Infor-

mation. PhD thesis, Saarland University, 2007.

Sebastian Pado and Mirella Lapata. Cross-lingual annotation projection of semantic

roles. Journal of Artificial Intelligence Research, 36:307–340, 2009.

Martha Palmer, Daniel Gildea, and Paul Kingsbury. The Proposition Bank: An anno-

tated corpus of semantic roles. Computational Linguistics, 31(1):71–105, 2005a.

Martha Palmer, Nianwen Xue, Olga Babko-Malaya, Jinying Chen, and Benjamin Sny-

der. A parallel Proposition Bank II for Chinese and English. In Proceedings of the

Workshop on Frontiers in Corpus Annotations II: Pie in the Sky, pages 61–67, Ann

Arbor, Michigan, June 2005b. Association for Computational Linguistics.

Martha Palmer, Dan Gildea, and Nianwen Xue. Semantic role labeling. Morgan &

Claypool Publishers, 2010.

Feng Pan, Rutu Mulkar-Mehta, and Jerry R. Hobbs. Annotating and learning event

durations in text. Computational Linguistics, 37(4):727–753, 2011.

David Pesetsky. Zero syntax: Experiencers and cascades. MIT Press, Cambridge Mass.,

1995.

284

Bibliography

James Pustejovsky. The generative lexicon. MIT Press, Cambridge, MA, 1995.

James Pustejovsky, Patrik Hanks, Roser Saurı, Andrew See, Robert Gaizauskas, Andrea

Setzer, Dragomir R. Radev, Beth Sundheim, David Day, Lisa Ferro, and Marzia Lazo.

The TIMEBANK corpus. In Corpus Linguistics, page 647–656, 2003.

Andrew Radford. Minimalist Syntax. Cambridge University Press, Cambridge, 2004.

Gillian Ramchand. Verb Meaning and the Lexicon: A First Phase Syntax. Cambridge

Studies in Linguistics. Cambridge University Press, Cambridge, 2008.

Tanja Reinhart. The theta system — An overview. Theoretical linguistics, 28:229–290,

2002.

Douglas LT Rohde. Tgrep2 user manual, 2004. URL http://tedlab.mit.edu/~dr/

Tgrep2/tgrep2.pdf.

Eleanor Rosch. Natural categories. Cognitive Psychology, 4(3):328–350, 1973.

Michael Roth and Anette Frank. Aligning predicate argument structures in monolin-

gual comparable texts: A new corpus for a new task. In *SEM 2012: The First Joint

Conference on Lexical and Computational Semantics – Volume 1: Proceedings of the

main conference and the shared task, and Volume 2: Proceedings of the Sixth Interna-

tional Workshop on Semantic Evaluation (SemEval 2012), pages 218–227, Montreal,

Canada, 7-8 June 2012. Association for Computational Linguistics.

Susan Rothstein. Telicity and atomicity. In Susan Rothstein, editor, Theoretical and

crosslinguistic approaches to the semantics of aspect, pages 43–78, Amsterdam, 2008.

John Benjamins.

Josef Ruppenhofer, Michael Ellsworth, Miriam R. L. Petruck, Christopher R. Johnson,

and Jan Scheffczyk. FrameNet II: Extended theory and practice, 2005. URL http:

//framenet.icsi.berkeley.edu/book/book.pdf.

Stuart J. Russell and Peter Norvig. Artificial intelligence : a modern approach. Prentice

Hall Pearson, Upper Saddle River, N.J., 2010.

285

Bibliography

Tanja Samardzic. Light verb constructions in English and Serbian. In English Language

and Literature Studies – Structures across Cultures, pages 59–73, Belgrade, 2008.

Faculty of Philology.

Tanja Samardzic, Lonneke van der Plas, Goldjihan Kashaeva, and Paola Merlo. The

scope and the sources of variation in verbal predicates in English and French. In

Markus Dickinson, Kaili Muurisep, and Marco Passarotti, editors, Proceedings of the

Ninth International Workshop on Treebanks and Linguistic Theories, volume 9, pages

199–211, Tartu, Estonia, 2010. Northern European Association for Language Tech-

nology (NEALT).

Tanja Samardzic and Paola Merlo. The meaning of lexical causatives in cross-linguistic

variation. Linguistic Issues in Language Technology, 7(12):1–14, 2012.

Florian Schafer. The causative alternation. In Language and Linguistics Compass,

volume 3, pages 641–681. Blackwell Publishing, 2009.

Sabine Schulte im Walde. Experiments on the choice of features for learning verb classes.

In Proceedings of The 10th Conference of the European Chapter of the Association for

Computational Linguistics (EACL 2003), pages 315–322, Budapest, Hungary, 2003.


Sabine Schulte im Walde. Experiments on the automatic induction of German semantic

verb classes. Computational Linguistics, 32(2):159–194, 2006.

Sabine Schulte im Walde, Christian Hying, Christian Scheible, and Helmut Schmid.

Combining EM training and the MDL principle for an automatic verb classification

incorporating selectional preferences. In Proceedings of ACL-08: HLT, pages 496–

504, Columbus, Ohio, June 2008. Association for Computational Linguistics. URL

http://www.aclweb.org/anthology/P/P08/P08-1057.

Violeta Seretan. Syntax-Based Collocation Extraction. Text, Speech and Language

Technology. Springer, Dordrecht, 2011.

Eric V. Siegel and Kathleen R. McKeown. Learning methods to combine linguistic

indicators: improving aspectual classification and revealing linguistic insights. Com-

putational Linguistics, 26(4):595–628, 2000.

286

Bibliography

Nate Silver. The Signal and the Noise: Why So Many Predictions Fail — but Some

Don’t. The Penguin Press, New York, 2012.

Benjamin Snyder, Tahira Naseem, Jacob Eisenstein, and Regina Barzilay. Unsupervised

multilingual learning for pos tagging. In Proceedings of the 2008 Conference on Em-

pirical Methods in Natural Language Processing, pages 1041–1050, Honolulu, 2008.


Benjamin Snyder, Tahira Naseem, and Regina Barzilay. Unsupervised multilingual

grammar induction. In Proceedings of the Joint Conference of the 47th Annual Meeting

of the ACL and the 4th International Joint Conference on Natural Language Process-

ing of the AFNLP, pages 73–81, Suntec, Singapore, August 2009. Association for


Maria Stambolieva. Parallel corpora in aspectual studies of non-aspect languages. In

Proceedings of The Second Workshop on Annotation and Exploitation of Parallel Cor-

pora, pages 39–42, Hissar, Bulgaria, September 2011.

Suzanne Stevenson and Eric Joanis. Semi-supervised verb class discovery using noisy

features. In Walter Daelemans and Miles Osborne, editors, Proceedings of the Seventh

Conference on Natural Language Learning at HLT-NAACL 2003, pages 71–78, 2003.

URL http://www.aclweb.org/anthology/W03-0410.pdf.

Suzanne Stevenson, Afsaneh Fazly, and Ryan North. Statistical measures of the semipro-

ductivity of light verb constructions. In Proceedings of the ACL’04 Workshop on Mul-

tiword Expressions: Integrating Processing, pages 1–8. Association for Computational

Linguistics, 2004.

Lin Sun and Anna Korhonen. Improving verb clustering with automatically acquired

selectional preferences. In Proceedings of the 2009 Conference on Empirical Methods

in Natural Language Processing, pages 638–647, Singapore, August 2009. Association

for Computational Linguistics. URL http://www.aclweb.org/anthology/D/D09/

D09-1067.

Peter Svenonius. Slavic prefixes inside and outside VP. Nordlyd, 32(2):205–253, 2004b.

Leonard Talmy. Towards a cognitive semantics. The MIT Press, Cambridge Mass., 2000.

287

Bibliography

Pasi Tapanainen, Jussi Piitulainen, and Timo Jarvinen. Idiomatic object usage and sup-

port verbs. In Proceedings of the 36th Annual Meeting of the Association for Compu-

tational Linguistics and 17th International Conference on Computational Linguistics,

Volume 2, pages 289–1293, Montreal, Quebec, Canada, August 1998. Association for


Jorg Tiedemann. News from OPUS - A collection of multilingual parallel corpora with

tools and interfaces. In Nicolas Nicolov, Kalina Bontcheva, Galia Angelova, and Ruslan

Mitkov, editors, Recent Advances in Natural Language Processing, volume V, pages

237–248, Borovets, Bulgaria, 2009. John Benjamins, Amsterdam/Philadelphia.

Ivan Titov and James Henderson. Constituent parsing with incremental sigmoid belief

networks. In Proceedings of the 45th Annual Meeting of the Association of Compu-

tational Linguistics, pages 632–639, Prague, Czech Republic, June 2007. Association

for Computational Linguistics.

Ivan Titov and Alexandre Klementiev. A Bayesian approach to unsupervised seman-

tic role induction. In Proceedings of the 13th Conference of the European Chapter of

the Association for Computational Linguistics, pages 12–22, Avignon, France, April

2012. Association for Computational Linguistics. URL http://www.aclweb.org/

anthology/E12-1003.

Kristina Toutanova, Aria Haghighi, and Christopher Manning. Joint learning improves

semantic role labeling. In Proceedings of the 43rd Annual Meeting of the Associa-

tion for Computational Linguistics (ACL’05), pages 589–596, Ann Arbor, Michigan,

June 2005. Association for Computational Linguistics. URL http://www.aclweb.

org/anthology/P/P05/P05-1073.

Graham Upton and Ian Cook. Understanding statistics. Oxford University Press, Oxford,

1996.

Lonneke van der Plas and Jorg Tiedemann. Finding synonyms using automatic word

alignment and measures of distributional similarity. In Proceedings of the COL-

ING/ACL 2006 Main Conference Poster Sessions, pages 866–873, Sydney, Australia,

July 2006. Association for Computational Linguistics.

288

Bibliography

Lonneke van der Plas, Tanja Samardzic, and Paola Merlo. Cross-lingual validity of

PropBank in the manual annotation of French. In Proceedings of the Fourth Lin-

guistic Annotation Workshop, pages 113–117, Uppsala, Sweden, 2010. Association for


Lonneke van der Plas, Paola Merlo, and James Henderson. Scaling up automatic cross-

lingual semantic role annotation. In Proceedings of the 49th Annual Meeting of the As-

sociation for Computational Linguistics: Human Language Technologies, pages 299–

304, Portland, Oregon, USA, June 2011. Association for Computational Linguistics.

Zeno Vendler. Linguistics in Philosophy. Cornell University Press, Ithaca, 1967.

Ruprecht von Waldenfels. Aspect in the imperative across Slavic - a corpus driven pilot

study. Oslo Studies in Language, 4(1):141–155, 2012.

Rok Zaucer. The reflexive-introducing na- and the distinction between internal and ex-

ternal Slavic prefixes. In Anastasia Smirnova, Vedrana Mihalicek, and Lauren Ressue,

editors, Formal Studies in Slavic Linguistics, pages 54–102, Newcastle upon Tyne,

2010. Cambridge Scholars Publishing.

Anna Wierzbicka. Why can you have a drink when you can’t *have an eat? Language,

58(4):753–799, 1982.

Edwin Williams. Lexical and synatctic complex predicates. In Alex Alsina, Joan Bres-

nan, and Peter Sells, editors, Complex predicates, pages 13–29, Stanford, California,

1997. CSLI Publications.

Jennifer Williams and Graham Katz. Extracting and modeling durations for habits and

events from twitter. In Proceedings of the 50th Annual Meeting of the Association

for Computational Linguistics (Volume 2: Short Papers), pages 223–227, Jeju Island,

Korea, July 2012. Association for Computational Linguistics.

Deirdre Wilson and Dan Sperber. Pragmatics and time. In Robyn Carston and Seiji

Uchida, editors, Relevance Theory: Applications and Implications, Amsterdam, 1998.

John Benjamins.

289

Bibliography

Ian H. Witten and Eibe Frank. Data mining : practical machine learning tools and

techniques. Morgan Kaufmann Publishers, San Francisco, 2005.

Phillip Wolff and Tatyana Ventura. When Russians learn English: How the semantics of

causation may change. Bilingualism: Language and Cognition, 12(2):153–176, 2009b.

Phillip Wolff, Ga-Hyun Jeon, and Yu Li. Causal agents in English, Korean and Chinese:

The role of internal and external causation. Language and Cognition, 1(2):165–194,

2009a.

Shumin Wu and Martha Palmer. Semantic mapping using automatic word alignment

and semantic role labeling. In Proceedings of Fifth Workshop on Syntax, Semantics

and Structure in Statistical Translation, pages 21–30, Portland, Oregon, USA, June

2011. Association for Computational Linguistics.

Nianwen Xue and Martha Palmer. Calibrating features for semantic role labeling. In

Dekang Lin and Dekai Wu, editors, Proceedings of Empirical Methods in Natural Lan-

guage Processing (EMNLP) 2004, pages 88–94, Barcelona, Spain, July 2004. Associ-

ation for Computational Linguistics.

David Yarowsky, Grace Ngai, and Richard Wicentowski. Inducing multilingual text

analysis tools via robust projection across aligned corpora. In Proceedings of the 1st

international conference Human Language Technology, pages 161–168, San Diego, CA,

2001. Association for Computational Linguistics.

Benat Zapirain, Eneko Agirre, and Lluıs Marquez. Robustness and generalization of

role sets: PropBank vs. VerbNet. In Proceedings of the Annual Meeting of the As-

sociation for Computational Linguistics (ACL) and the Human Language Technology

Conference, pages 550–558, Columbus, Ohio, 2008. Association for Computational

Linguistics. URL http://www.aclweb.org/anthology/P/P08/P08-1063.

Sina Zarrieß and Jonas Kuhn. Exploiting translational correspondences for pattern-

independent mwe identification. In Proceedings of the Workshop on Multiword Expres-

sions: Identification, Interpretation, Disambiguation and Applications, pages 23–30,

Singapore, August 2009. Association for Computational Linguistics.

290

Bibliography

Sina Zarrieß, Aoife Cahill, Jonas Kuhn, and Christian Rohrer. A cross-lingual induction

technique for german adverbial participles. In Proceedings of the 2010 Workshop on

NLP and Linguistics: Finding the Common Ground, pages 34–42, Uppsala, Sweden,

July 2010. Association for Computational Linguistics.

291

A. Light verb constructions data

A.1. Word alignment of the constructions with ’take’

Mapping EN DE Target DE Target EN

v n v n

2 – 2 taken account berucksichtigung finden no no bad good

2 – 1 take account berucksichtigen no good good good

2 – 1 take account berucksichtigen no no good good

2 – 1 taken action aktion no good bad good

2 – 2 take check kontrolle durchzufuhren no no bad bad

2 – 2 taken initiative initiative ergriffen no good good good

2 – 1 take precedence dominieren no no bad good

2 – 2 take decision beschluss fassen no good bad good

2 – 2 take account rechnung tragen no good bad good

2 – 2 take initiative initiative ergriffen no good good good

2 – 1 take seat sitzen no bad bad bad

2 – 2 approach taken kurs verfolgt no bad bad bad

2 – 1 decision taken beschlossen no bad good bad


2 – 2 take note nehmen (zur) kenntnis no good good good

2 – 1 initiative taken initiative no good bad good

2 – 2 take view sind (der) ansicht no good bad good


2 – 1 take account berucksichtigt no no good good

293


2 – 2 take action maßnahmen ergreifen good good good good

2 – 2 taken view meinung vertreten no good good good

2 – 1 taken initiative vorantreiben no no bad bad

2 – 2 take view vetreten meinung no good good good

2 – 1 take action aufarbeitet no no good good

2 – 2 took account rechnung tragen no no good good

2 – 2 took steps schritte unternahme no good good good

2 – 2 taken note (zur) kenntnis genommen no no good good

2 – 1 taken initiative initiative no good bad bad

2 – 2 take view bin (der) ansicht no good bad good

2 – 2 take steps schritte unternehmen no good bad good

2 – 1 take decision entscheiden no no good good

2 – 1 take cognisance berucksichtigen no bad bad bad

2 – 1 take account berucksichtigen no good good good

2 – 2 decision taken entscheidung getroffen good good good good

2 – 1 decision taken entscheiden no no good good

2 – 0 actions take no translation no no bad bad

2 – 2 taken step schritt kommen no good bad good

2 – 1 steps taken schritte no good bad good

2 – 1 take decisions beschließen no no good good

2 – 1 take care kummert no no good good

2 – 2 decision taken beschlusse fassen no good bad bad

2 – 2 taken note (zur) kenntnis genommen no good good good

2 – 0 taken decisions no translation no no bad bad

2 – 2 take action maßnahmen ergriffen no good good good

2 – 2 take steps schritte unternehmen no good bad good

2 – 2 decision taken gefaßten beschlussen no good good good


2 – 2 decision taken getroffenen entscheidung good good good good

2 – 2 steps taken schritte vollziehen no good bad good

2 – 1 take decision entscheidung no good bad good

2 – 1 take approach herantreten no bad bad bad

2 – 2 take break pause einlegen no good good good

294

A.1. Word alignment of the constructions with ’take’

2 – 1 take (into) account berucksichtigt no good good good

2 – 2 action taken sanktionen verhangt good bad good good

2 – 2 steps taken maßnahmen ergriffen no good good good

2 – 2 take decision entscheidung treffen no good good good

2 – 1 take action einschreiten no no bad bad

2 – 1 take notice berucksichtigen no no good good

2 – 2 take view sind (der) ansicht no good bad good

2 – 2 decisions taken beschlusse gefaßt no good good good

2 – 1 take (into) account ubernehmen no no good bad


2 – 2 take view (um) standpunkt vertreten no no bad good

2 – 1 taken (into) account berucksichtigt no no good good

2 – 2 decision taken beschluß angenommen no good good good

2 – 1 taken account berucksichtigt no good good good

2 – 2 decision taken beschlusse getroffen no good good good

2 – 2 steps taken hurden nehmen no no bad bad

2 – 2 steps take schritte unternimmt bad good good good

2 – 2 took view ansicht vertreten bad good good good

2 – 2 took decision beschluß gefaßt bad good good good

2 – 1 approach took vorgehen no no bad bad

2 – 1 took account erortert bad no good good

2 – 2 take steps maßnahmen ergriffen no good good good

2 – 2 decisions taken entscheidungen getroffen no good good good

2 – 2 vote taken abstimmung findet-statt no good good good

2 – 1 take (into) account berucksichtigen no good good good

2 – 2 samples taken proben gezogen no good bad good

2 – 1 take look anschaut no good bad good

2 – 2 obligations take verpflichtungen

wahrnehmen

no good good good

2 – 1 takes account berucksichtigt no good good good

2 – 2 decision taken getroffene entscheidung no good good good

2 – 1 action taken maßnahmen no good bad bad

2 – 0 action taken no translation no no bad bad

295


2 – 2 take decisions entscheidungen treffen no good good good

2 – 1 take (into) account einbezogen no no bad bad

2 – 1 taken note notiert no good good good

2 – 2 decisions taken getroffene beschlusse no no good good

2 – 2 take notice berucksichtigen no no bad bad

2 – 2 took step schritt vollzogen no good bad good

2 – 2 take control kontrolle bringen no good bad good

2 – 0 step take no translation bad bad bad bad

2 – 1 taken (into) account berucksichtigt no good good good

2 – 0 action taken no translation no no bad bad


2 – 2 decisions taken getroffene entscheidungen no good bad good

2 – 2 decisions taken gefallten entscheidungen no good bad good


2 – 2 took decision beschluß gefaßt bad good good good

2 – 1 decisions taken beschlusse bad good good good

296

A.2. Word alignment of the constructions with ’make’


Mapping EN DE Target DE Target EN

v n v n

2 – 1 make choices auszuwahlen no no good good

2 – 0 make use no translation no no bad bad

2 – 1 make progress vorankommen no no good good

2 – 2 makes cuts kurzungen vornimmt no no bad bad

2 – 2 make decisions entscheiden betreffen no no bad bad

2 – 1 make contribution beitragen no bad bad bad

2 – 2 make decisions entscheidungen treffen no good good good

2 – 1 make reduction reduziert no no bad bad

2 – 1 make start anfangen no no bad bad

2 – 2 make points punkte ansprechen no good good good

2 – 1 make use einsetzen no no bad bad

2 – 2 make point verfahrensfrage anzus-

prechen

no good good good

2 – 2 make contribution beitrag geleistet no good bad good

2 – 1 speech made rede no good bad bad

2 – 2 comparison made vergleich anstellen no good good good

2 – 1 investments made investiert no bad bad bad

2 – 2 progress made fortschritte erzielt no good good good

2 – 2 make comments bemerkungen machen no good good good

2 – 1 make decisions entscheiden no no good good

2 – 1 make comment sagen no good good good

2 – 1 make provision einrichten no no bad bad

2 – 1 comments make anmerkungen no good good good

2 – 1 proposal makes vorschlag no good bad good

2 – 2 make suggestion anmerkungen machen no no good bad

2 – 2 make checks prufungen vornehmen no no good bad

2 – 2 make progress schritt getan no no bad bad

2 – 1 make changes korrekturen no no bad good

297


2 – 1 comments made aussagen no good good good

2 – 1 choice made entscheidung bad good bad bad

2 – 1 changes made geandert no no good good

2 – 1 attempts made versucht no good bad bad

2 – 2 made request forderung erhoben no bad good bad

2 – 1 attempts made versucht no good good bad

2 – 1 investments made investitionen no good bad good

2 – 1 points made gesagt no no bad bad

2 – 2 made contribution beitrag geleistet no good good good

2 – 1 investments made investiert no no bad bad

2 – 2 gains made erzielten erfolge no no good bad

2 – 2 made progress kommt voran no no bad good

2 – 0 makes fuss no translation no no bad bad

2 – 1 makes reference befaßt no no bad bad

2 – 2 make statement erklarung abgegeben bad good bad bad

2 – 1 make inspections kontrollen bad good bad bad

2 – 2 make assessment bilanz ziehen no bad bad bad

2 – 1 make statement (um) wort no no bad bad

2 – 1 make remarks eingehen no no bad bad

2 – 2 make observation bemerkung machen no good good bad

2 – 2 made contribution beitrag leisten no good good good

2 – 2 made start hat start no no bad good

2 – 1 suggestion made anregung no good good good

2 – 1 mistakes made fehler no good good good

2 – 1 comment made anmerkung bad good good good

2 – 2 progress made fortschritte erzielt bad good good good

2 – 1 reference made erwahnt bad no bad bad

2 – 1 appeal made aufruf no good good good

2 – 1 references made verweis no good good good

2 – 2 appointments made einstellungen vorgenom-

men

no good good good

2 – 2 progress made fortschritte erzielt no good bad good

298


2 – 2 decisions made entscheidungen

berucksichtigt

no good bad good

2 – 0 reference made no translation no no bad bad

2 – 2 decisions made entscheidung fallen good good good good

2 – 2 statement made abgegebenen erklarung good good good good

2 – 1 comments made bemerkungen no good good good

2 – 1 made statement ausgesagt no no good good

2 – 2 make contribution beitrag liefern no good good good

2 – 2 make proposal vorschlag machen no good good good

2 – 2 make reference bezug nehmen no no bad bad

2 – 2 make contribution beitrag leisten no good good good

2 – 2 make progress fortschritte erzielen no good bad good

2 – 2 make contribution beitrag leisten good good bad good

2 – 1 make assessment einschatzungsvermogen no bad good bad

2 – 0 make point no translation no no bad bad

2 – 1 make demands uberfordern no no good good

2 – 2 make statement erklarung abgeben no good bad good

2 – 2 contribution make beitrag leisten no good good good

2 – 2 make use gebrauch machen good no good good

2 – 2 make contribution beitrag aufgaben no good bad good

2 – 2 make changes sehen veranderungen no no bad good

2 – 2 make contribution beitrag leisten no good good good

2 – 2 make decisions macht haben no no bad bad

2 – 1 make points bemerkungen no good bad good

2 – 1 make profits verdienstmoglichkeiten no bad good good

2 – 2 achievements made erreichten erfolge no good bad bad

2 – 2 made proposal vorschlag gelesen no good bad good

2 – 2 points made punkte angesprochen no good good good

2 – 2 made attempts versuch unternommen no no good bad

2 – 1 points made punkte no good bad bad

2 – 2 demands made forderungen gestellt good good bad good

2 – 1 calls made gefordert no good bad bad

2 – 2 made proposal vorschlag gemacht no good good good

299


2 – 1 made decision entschieden no bad bad bad

2 – 2 decisions made entscheidungen no good good good

2 – 1 made pronounce-

ments

gesagt no good bad bad

2 – 2 made comment bemerkung gemacht good good good good

2 – 1 progress made fortschritte no good good good

2 – 2 proposal made vorschlag machen no good bad good

2 – 1 promises made versprechen no good good good

2 – 1 attempt made versucht no good bad bad

2 – 1 use made forderung no no bad bad

2 – 2 makes changes anderungen vorgeshclagen no good bad good

300

A.3. Word alignments of regular constructions


2 – 2 create basis grundlage schaffen good good good good

2 – 1 jobs created arbeitsplatze bad good bad good

2 – 2 created climate klima schaffen no good bad good

2 – 2 create framework entsteht rahmen no good good good

2 – 2 jobs created arbeitsplatze geschaffen good good good good

2 – 2 create regime regelung schaffen no good good good

2 – 2 create inequality ware ungleichheit no good bad good

2 – 0 create tape no translation no bad bad bad

2 – 2 creates networks verwirklichung

verkehrsnetze

no good bad good

2 – 2 they created sie dimensioniert good good good good

2 – 2 create area finanzraum schaffen good no good good

2 – 2 create jobs schafft arbeitsplatze good good bad good

2 – 2 jobs created schaffung arbeitsplatzen good good bad good

2 – 2 create problems probleme herauf-

beschworen

no good good good

2 – 2 create inequalities ungleichheit schafft good good bad good

2 – 2 consensus created war einig no no bad good

2 – 2 create jobs arbeitsplatze schaffen good good good good

2 – 2 create incentives anreize schaffen good good good bad

2 – 2 create institution institutionen schaffen no good good bad

2 – 2 peace created frieden schaffen good good good good

2 – 2 create charter titel einfugen no good good bad

2 – 2 create conditions beitrittsfahigkeit

herzustellen

no no good good

2 – 2 create council sicherheitsrat schaffen good bad good good

2 – 2 jobs created arbeitsplatze entstehen no good bad good

2 – 2 create societies formierung gesellschaft bad no good bad

2 – 2 jobs created arbeitsplatze geschaffen good good good good

2 – 2 create problem schafft problem no no good good

2 – 2 create code verhaltenskodex schaffen good good good good

301


2 – 2 create union union vereinbaren no no good good

2 – 2 create source aufbau informationssys-

tems

no no good good

2 – 2 literature produced produzierten litaratur no good bad good

2 – 2 food produced nahrungsmittel produziert good good good good

2 – 2 produce cereals anbau qualitatsgetreide no no bad good

2 – 2 produce paper grunbuch vorgelegt good no bad good

2 – 2 produce industry entwicklung indus-

triezweigs

no no bad good

2 – 2 produce programme vorbereitung programms no good bad good

2 – 2 food produce produzierten nahrungsmit-

tel

good good bad good

2 – 2 produce wine wein erzeugen no good good good

2 – 2 sherry produced hergestellten sherry no good good good

2 – 2 draw outlines rahmen vorgegeben no good bad bad

2 – 2 parities fixed paritaten festgelegt no good bad good

2 – 2 constructed europe europa aufgebau no no bad good

2 – 2 reconstruct kosovo kosovo wiederaufbauen no good good good

2 – 2 reconstruct balkans wiederaufbau balkan good good bad good

2 – 2 rebuild confidence vertrauen aufbauen good good good good

2 – 2 build bureaucracy verwaltungsapparat auf-

bauen

no good good good

2 – 2 build democracy aufbau demokratie no no bad bad

2 – 2 establish shelter bereitstellung un-

terkunften

no no bad good

2 – 2 commission estab-

lished

kommission eingesetzt good good bad good

2 – 2 establish framework rahmen setzen no good bad good

2 – 2 priorities estab-

lished

festgeschriebenen pri-

oritaten

no good good good

2 – 2 policy established preispolitik stabilisiert bad good good good

2 – 2 establish norms arbeitsnormen einsetzen no bad bad good

2 – 0 establish principle no translation no no bad bad

302


2 – 2 establish consis-

tency

konsequenz verstarkt no good bad good

2 – 2 establish systems identifizierungssysteme

festgelegt

bad no bad good

2 – 2 establish vanguard bildung vorhut no good good good

2 – 2 primacy established primat herausgestellt no no good good

2 – 2 multinationals

established

konzerne niedergelassen no no bad bad

2 – 2 establish founda-

tions

grundlagen schaffen good good good good

2 – 2 establish policy umweltpolitik machen no bad bad good

2 – 2 establish clarity klarheit schaffen no good bad good

2 – 2 they established sie aufgebaut good good good good

2 – 2 establish procedures verfahren hervorgebracht no good bad good

2 – 2 established stages phasen festgelegt good good bad good

2 – 2 criteria established aufgestellten kriterien no good good good

2 – 2 established perspec-

tives

vorausschau aufgestellt bad good bad good

2 – 2 procedures estab-

lished

vereinbarenden ver-

fahrensweiseg

good no good good

2 – 2 establish conditions bedingungen schafft no good bad good

2 – 2 partnerships estab-

lished

beitrittspartnerschaft

besteht

no bad bad good

2 – 2 establish itself sich festigen no no bad bad

2 – 2 create situation situation herauf-

beschworen

bad good good good

2 – 2 create peace schaffung friedens good good good good

2 – 2 created alternatives alternativen geschaffen good good good good

2 – 2 sherry produced hergestellten sherry no good good good

2 – 2 establish system system schaffen good good bad good

2 – 2 establish court satzung strafgerichtshofs no good good good

2 – 2 create fund schaffung fonds good good good good

303


2 – 2 opportunities cre-

ated

moglichkeiten bietet no good bad good

2 – 2 created instruments gibt instrument no good bad good

2 – 2 create opportunity moglichkeit finden no good bad good

2 – 2 opportunities cre-

ated

arbeitsmoglichkeiten

geschaffen

good good good good

2 – 2 create sources bauen spannungsfaktoren no no bad good

2 – 2 create problems probleme verursachen no good good good

2 – 2 jobs create arbeitsplatze geschaffen good good good good

2 – 2 create opportunities schaffung arbeitsplatze good good bad bad

2 – 2 create conditions bedingungen schaffen no good good good

2 – 2 created policy wirtschaftspolitik verwirk-

licht

no no good good

2 – 2 produce products produktion liefern bad no bad bad

2 – 1 produce goods produktion good no bad bad

2 – 2 produce them sie herstellen no good bad bad

2 – 2 produce obstacles handelshemmnisse erzeu-

gen

no good bad good

2 – 2 establish conditions festlegung einstellungsbe-

dingungen

no no bad bad

2 – 2 establish priorities sind prioritaten no no bad bad

2 – 1 establishes right legt good no good bad

2 – 2 establish partner-

ships

regionalpartnerschaften

einzugehen

no good bad good

2 – 1 establish democracy demokratischen no good bad bad

2 – 2 distance established mindestentfernung einge-

halten

no no bad good

2 – 1 establish chapter charta no good bad good

2 – 2 this established das festlegen no good good good

304

B. Corpus counts and measures for

lexical causatives

Verb Counts Caus.

rate

Anticaus.

rate

Passive

rate

C/A ra-

tio

Sp-value

abate 11 0.18 0.82 0 0.22 -1.5

accelerate 128 0.52 0.22 0.27 2.36 0.86

acetify 0 0.19 0.31 0.5 0.63 -0.47

acidify 0 0.19 0.31 0.5 0.63 -0.47

age 63 0.13 0.57 0.3 0.22 -1.5

agglomerate 0 0.19 0.31 0.5 0.63 -0.47

air 40 0.18 0.05 0.78 3.5 1.25

alkalify 0 0.19 0.31 0.5 0.63 -0.47

alter 436 0.58 0.08 0.34 7.17 1.97

ameliorate 4 0.05 0.5 0.5 0.1 -2.34

americanize 0 0.19 0.31 0.5 0.63 -0.47

asphyxiate 0 0.19 0.31 0.5 0.63 -0.47

atrophy 2 0.5 0.5 0 1 0

attenuate 3 0.33 0.1 0.67 3.25 1.18

awaken 29 0.55 0.21 0.24 2.67 0.98

balance 238 0.18 0.05 0.78 3.82 1.34

beam 0 0.19 0.31 0.5 0.63 -0.47

beep 0 0.19 0.31 0.5 0.63 -0.47

bend 17 0.47 0.53 0 0.89 -0.12

bivouac 0 0.19 0.31 0.5 0.63 -0.47

305

B. Corpus counts and measures for lexical causatives

Verb Counts Caus.

rate

Anticaus.

rate

Passive

rate

C/A ra-

tio

Sp-value

blacken 4 1 0.08 0 13 2.56

blare 0 0.19 0.31 0.5 0.63 -0.47

blast 4 0.05 1 0 0.05 -3.03

bleed 20 0.25 0.65 0.1 0.38 -0.96

blink 0 0.19 0.31 0.5 0.63 -0.47

blunt 1 1 0.31 0 3.25 1.18

blur 23 0.43 0.43 0.13 1 0

board 9 0.56 0.11 0.33 5 1.61

bounce 12 0.17 0.75 0.08 0.22 -1.5

break 701 0.33 0.3 0.38 1.1 0.09

brighten 2 0.1 0.5 0.5 0.19 -1.65

broaden 110 0.36 0.14 0.5 2.67 0.98

brown 0 0.19 0.31 0.5 0.63 -0.47

burn 137 0.23 0.18 0.59 1.24 0.22

burp 0 0.19 0.31 0.5 0.63 -0.47

burst 14 0.29 0.71 0 0.4 -0.92

buzz 1 0.19 1 0 0.19 -1.65

calcify 0 0.19 0.31 0.5 0.63 -0.47

canter 0 0.19 0.31 0.5 0.63 -0.47

capsize 4 0.05 1 0 0.05 -3.03

caramelize 0 0.19 0.31 0.5 0.63 -0.47

carbonify 0 0.19 0.31 0.5 0.63 -0.47

carbonize 0 0.19 0.31 0.5 0.63 -0.47

change 3457 0.37 0.42 0.2 0.87 -0.14

char 0 0.19 0.31 0.5 0.63 -0.47

cheapen 1 1 0.31 0 3.25 1.18

cheer 10 0.2 0.4 0.4 0.5 -0.69

chill 1 0.19 1 0 0.19 -1.65

choke 9 0.22 0.44 0.33 0.5 -0.69

clack 0 0.19 0.31 0.5 0.63 -0.47

clang 0 0.19 0.31 0.5 0.63 -0.47

306

Verb Counts Caus.

rate

Anticaus.

rate

Passive

rate

C/A ra-

tio

Sp-value

clash 21 0.1 0.9 0 0.11 -2.25

clatter 0 0.19 0.31 0.5 0.63 -0.47

clean 58 0.53 0.12 0.34 4.43 1.49

clear 296 0.5 0.07 0.43 7.1 1.96

click 0 0.19 0.31 0.5 0.63 -0.47

clog 6 0.17 0.5 0.33 0.33 -1.1

close 1604 0.2 0.14 0.66 1.47 0.39

coagulate 0 0.19 0.31 0.5 0.63 -0.47

coarsen 0 0.19 0.31 0.5 0.63 -0.47

coil 0 0.19 0.31 0.5 0.63 -0.47

collapse 151 0.04 0.95 0.01 0.04 -3.18

collect 249 0.27 0.06 0.68 4.71 1.55

compress 3 0.33 0.1 0.67 3.25 1.18

condense 2 1 0.15 0 6.5 1.87

contracte 0 0.19 0.31 0.5 0.63 -0.47

cool 13 0.23 0.46 0.31 0.5 -0.69

corrode 2 0.5 0.5 0 1 0

crack 22 0.27 0.5 0.23 0.55 -0.61

crash 15 0.01 1 0 0.01 -4.36

crease 0 0.19 0.31 0.5 0.63 -0.47

crimson 0 0.19 0.31 0.5 0.63 -0.47

crinkle 0 0.19 0.31 0.5 0.63 -0.47

crisp 0 0.19 0.31 0.5 0.63 -0.47

crumble 35 0.03 0.94 0.03 0.03 -3.5

crumple 0 0.19 0.31 0.5 0.63 -0.47

crystallize 2 1 0.15 0 6.5 1.87

dampen 9 1 0.03 0 29.25 3.38

dangle 5 0.6 0.2 0.2 3 1.1

darken 4 0.5 0.08 0.5 6.5 1.87

decelerate 1 0.19 0.31 1 0.63 -0.47

decentralize 9 0.22 0.03 0.78 6.5 1.87

307


Verb Counts Caus.

rate

Anticaus.

rate

Passive

rate

C/A ra-

tio

Sp-value

decompose 0 0.19 0.31 0.5 0.63 -0.47

decrease 253 0.11 0.82 0.07 0.14 -1.97

deepen 84 0.42 0.39 0.19 1.06 0.06

deflate 3 0.06 0.67 0.33 0.1 -2.34

defrost 1 0.19 0.31 1 0.63 -0.47

degenerate 59 0 0.98 0.02 0 -5.71

degrade 20 0.15 0.05 0.8 3 1.1

dehumidify 0 0.19 0.31 0.5 0.63 -0.47

delight 834 0.09 0 0.91 35.5 3.57

demagnetize 0 0.19 0.31 0.5 0.63 -0.47

democratize 4 0.25 0.08 0.75 3.25 1.18

depressurize 0 0.19 0.31 0.5 0.63 -0.47

desiccate 0 0.19 0.31 0.5 0.63 -0.47

destabilize 19 0.68 0.02 0.32 42.25 3.74

deteriorate 291 0.03 0.96 0.01 0.03 -3.43

detonate 5 0.2 0.06 0.8 3.25 1.18

dim 4 0.75 0.25 0 3 1.1

diminish 189 0.31 0.4 0.29 0.76 -0.27

dirty 0 0.19 0.31 0.5 0.63 -0.47

disintegrate 14 0.14 0.86 0 0.17 -1.79

dissipate 16 0.44 0.25 0.31 1.75 0.56

dissolve 63 0.17 0.11 0.71 1.57 0.45

distend 0 0.19 0.31 0.5 0.63 -0.47

divide 686 0.16 0.04 0.8 4.19 1.43

double 220 0.21 0.63 0.15 0.34 -1.08

drain 26 0.54 0.12 0.35 4.67 1.54

drift 46 0.13 0.87 0 0.15 -1.9

drive 586 0.36 0.11 0.53 3.33 1.2

drop 303 0.22 0.42 0.36 0.54 -0.62

drown 51 0.2 0.49 0.31 0.4 -0.92

dry 18 0.06 0.94 0 0.06 -2.83

308

Verb Counts Caus.

rate

Anticaus.

rate

Passive

rate

C/A ra-

tio

Sp-value

dull 2 1 0.15 0 6.5 1.87

ease 98 0.67 0.14 0.18 4.71 1.55

empty 27 0.22 0.22 0.56 1 0

emulsify 0 0.19 0.31 0.5 0.63 -0.47

energize 0 0.19 0.31 0.5 0.63 -0.47

enlarge 183 0.16 0.33 0.51 0.5 -0.69

enthuse 3 0.33 0.33 0.33 1 0

equalize 0 0.19 0.31 0.5 0.63 -0.47

evaporate 20 0.01 1 0 0.01 -4.64

even 9 0.44 0.22 0.33 2 0.69

expand 343 0.2 0.52 0.28 0.39 -0.94

explode 50 0.04 0.92 0.04 0.04 -3.14

fade 40 0.05 0.95 0 0.05 -2.94

fatten 4 0.05 0.08 1 0.63 -0.47

federate 1 1 0.31 0 3.25 1.18

fill 265 0.37 0.08 0.55 4.71 1.55

firm 3 0.33 0.1 0.67 3.25 1.18

flash 3 0.06 1 0 0.06 -2.75

flatten 7 0.29 0.14 0.57 2 0.69

float 34 0.38 0.18 0.44 2.17 0.77

flood 88 0.26 0.33 0.41 0.79 -0.23

fly 114 0.22 0.78 0 0.28 -1.27

fold 7 0.03 1 0 0.03 -3.59

fossilize 0 0.19 0.31 0.5 0.63 -0.47

fracture 4 0.25 0.08 0.75 3.25 1.18

fray 1 0.19 1 0 0.19 -1.65

freeze 115 0.18 0.04 0.77 4.2 1.44

freshen 2 1 0.15 0 6.5 1.87

frost 0 0.19 0.31 0.5 0.63 -0.47

fructify 0 0.19 0.31 0.5 0.63 -0.47

fuse 3 0.33 0.33 0.33 1 0

309


Verb Counts Caus.

rate

Anticaus.

rate

Passive

rate

C/A ra-

tio

Sp-value

gallop 5 0.2 0.8 0 0.25 -1.39

gasify 0 0.19 0.31 0.5 0.63 -0.47

gelatinize 0 0.19 0.31 0.5 0.63 -0.47

gladden 1 0.19 0.31 1 0.63 -0.47

glide 0 0.19 0.31 0.5 0.63 -0.47

glutenize 0 0.19 0.31 0.5 0.63 -0.47

granulate 0 0.19 0.31 0.5 0.63 -0.47

gray 0 0.19 0.31 0.5 0.63 -0.47

green 0 0.19 0.31 0.5 0.63 -0.47

grieve 6 0.17 0.67 0.17 0.25 -1.39

grow 1379 0.14 0.78 0.08 0.19 -1.68

halt 130 0.35 0.05 0.61 7.5 2.01

hang 106 0.25 0.62 0.12 0.41 -0.89

harden 14 0.21 0.64 0.14 0.33 -1.1

harmonize 89 0.28 0.07 0.65 4.17 1.43

hasten 23 0.3 0.65 0.04 0.47 -0.76

heal 31 0.29 0.39 0.32 0.75 -0.29

heat 26 0.27 0.27 0.46 1 0

heighten 59 0.53 0.07 0.41 7.75 2.05

hoot 0 0.19 0.31 0.5 0.63 -0.47

humidify 0 0.19 0.31 0.5 0.63 -0.47

hush 18 0.01 0.11 0.89 0.1 -2.34

hybridize 0 0.19 0.31 0.5 0.63 -0.47

ignite 6 0.83 0.17 0 5 1.61

improve 3021 0.45 0.22 0.33 2.03 0.71

increase 4292 0.39 0.41 0.2 0.95 -0.05

incubate 4 0.5 0.5 0 1 0

inflate 12 0.42 0.08 0.5 5 1.61

intensify 242 0.46 0.24 0.31 1.95 0.67

iodize 0 0.19 0.31 0.5 0.63 -0.47

ionize 2 0.1 0.15 1 0.63 -0.47

310

Verb Counts Caus.

rate

Anticaus.

rate

Passive

rate

C/A ra-

tio

Sp-value

jangle 0 0.19 0.31 0.5 0.63 -0.47

jingle 0 0.19 0.31 0.5 0.63 -0.47

jump 61 0.39 0.61 0 0.65 -0.43

kindle 9 0.33 0.03 0.67 9.75 2.28

lean 24 0.04 0.92 0.04 0.05 -3.09

leap 11 0.02 1 0 0.02 -4.05

lengthen 19 0.42 0.32 0.26 1.33 0.29

lessen 51 0.61 0.2 0.2 3.1 1.13

level 145 0.18 0.16 0.66 1.13 0.12

levitate 0 0.19 0.31 0.5 0.63 -0.47

light 28 0.36 0.18 0.46 2 0.69

lighten 17 0.71 0.12 0.18 6 1.79

lignify 0 0.19 0.31 0.5 0.63 -0.47

liquefy 0 0.19 0.31 0.5 0.63 -0.47

lodge 115 0.39 0.02 0.59 22.5 3.11

loop 0 0.19 0.31 0.5 0.63 -0.47

loose 2 0.5 0.15 0.5 3.25 1.18

loosen 8 0.5 0.04 0.5 13 2.56

macerate 0 0.19 0.31 0.5 0.63 -0.47

madden 0 0.19 0.31 0.5 0.63 -0.47

magnetize 0 0.19 0.31 0.5 0.63 -0.47

magnify 7 0.14 0.04 0.86 3.25 1.18

march 25 0.08 0.88 0.04 0.09 -2.4

mature 30 0.01 0.9 0.1 0.01 -4.94

mellow 0 0.19 0.31 0.5 0.63 -0.47

melt 24 0.01 0.92 0.08 0.01 -4.74

moisten 0 0.19 0.31 0.5 0.63 -0.47

move 2910 0.11 0.8 0.09 0.14 -1.97

muddy 2 0.5 0.15 0.5 3.25 1.18

multiply 102 0.27 0.49 0.24 0.56 -0.58

narrow 56 0.36 0.34 0.3 1.05 0.05

311


Verb Counts Caus.

rate

Anticaus.

rate

Passive

rate

C/A ra-

tio

Sp-value

neaten 0 0.19 0.31 0.5 0.63 -0.47

neutralize 4 0.25 0.08 0.75 3.25 1.18

nitrify 0 0.19 0.31 0.5 0.63 -0.47

obsess 26 0.04 0.01 0.96 3.25 1.18

open 1627 0.54 0.14 0.32 3.79 1.33

operate 994 0.1 0.84 0.06 0.12 -2.16

ossify 1 0.19 0.31 1 0.63 -0.47

overturn 78 0.29 0.01 0.69 23 3.14

oxidize 0 0.19 0.31 0.5 0.63 -0.47

pale 1 0.19 1 0 0.19 -1.65

perch 1 0.19 0.31 1 0.63 -0.47

petrify 1 0.19 0.31 1 0.63 -0.47

polarize 0 0.19 0.31 0.5 0.63 -0.47

pop 4 0.25 0.75 0 0.33 -1.1

proliferate 20 0.15 0.85 0 0.18 -1.73

propagate 13 0.38 0.08 0.54 5 1.61

purify 1 0.19 0.31 1 0.63 -0.47

purple 0 0.19 0.31 0.5 0.63 -0.47

putrefy 1 0.19 1 0 0.19 -1.65

puzzle 30 0.17 0.07 0.77 2.5 0.92

quadruple 10 0.4 0.6 0 0.67 -0.41

quicken 3 0.67 0.33 0 2 0.69

quiet 1 0.19 1 0 0.19 -1.65

quieten 2 0.1 1 0 0.1 -2.34

race 6 0.17 0.5 0.33 0.33 -1.1

redden 0 0.19 0.31 0.5 0.63 -0.47

regularize 3 0.06 0.33 0.67 0.19 -1.65

rekindle 13 0.54 0.08 0.38 7 1.95

reopen 95 0.45 0.08 0.46 5.38 1.68

reproduce 47 0.3 0.3 0.4 1 0

rest 258 0.05 0.94 0 0.06 -2.85

312

Verb Counts Caus.

rate

Anticaus.

rate

Passive

rate

C/A ra-

tio

Sp-value

revolve 22 0.32 0.68 0 0.47 -0.76

ring 50 0.28 0.54 0.18 0.52 -0.66

rip 12 0.33 0.03 0.67 13 2.56

ripen 1 0.19 0.31 1 0.63 -0.47

roll 38 0.37 0.16 0.47 2.33 0.85

rotate 5 0.04 0.6 0.4 0.06 -2.75

roughen 0 0.19 0.31 0.5 0.63 -0.47

round 67 0.66 0.07 0.27 8.8 2.17

rumple 0 0.19 0.31 0.5 0.63 -0.47

run 1293 0.3 0.56 0.14 0.53 -0.64

rupture 1 0.19 0.31 1 0.63 -0.47

rustle 0 0.19 0.31 0.5 0.63 -0.47

sadden 59 0.17 0.05 0.78 3.33 1.2

scorch 2 0.1 0.15 1 0.63 -0.47

sear 0 0.19 0.31 0.5 0.63 -0.47

settle 455 0.17 0.21 0.62 0.83 -0.18

sharpen 13 0.69 0.15 0.15 4.5 1.5

shatter 52 0.29 0.06 0.65 5 1.61

shelter 17 0.47 0.24 0.29 2 0.69

shine 18 0.17 0.83 0 0.2 -1.61

short 0 0.19 0.31 0.5 0.63 -0.47

short- 0 0.19 0.31 0.5 0.63 -0.47

shorten 55 0.25 0.07 0.67 3.5 1.25

shrink 80 0.08 0.93 0 0.08 -2.51

shrivel 0 0.19 0.31 0.5 0.63 -0.47

shut 155 0.3 0.1 0.61 3.07 1.12

sicken 3 0.33 0.1 0.67 3.25 1.18

silicify 0 0.19 0.31 0.5 0.63 -0.47

silver 0 0.19 0.31 0.5 0.63 -0.47

singe 0 0.19 0.31 0.5 0.63 -0.47

sink 120 0.1 0.82 0.08 0.12 -2.1

313


Verb Counts Caus.

rate

Anticaus.

rate

Passive

rate

C/A ra-

tio

Sp-value

sit 723 0.06 0.93 0.01 0.07 -2.71

slack 0 0.19 0.31 0.5 0.63 -0.47

slacken 12 0.42 0.58 0 0.71 -0.34

slide 27 0.11 0.89 0 0.13 -2.08

slim 4 0.05 0.25 0.75 0.19 -1.65

slow 156 0.48 0.38 0.13 1.25 0.22

smarten 0 0.19 0.31 0.5 0.63 -0.47

smooth 14 0.64 0.02 0.36 29.25 3.38

snap 4 0.25 0.5 0.25 0.5 -0.69

soak 4 0.05 0.25 0.75 0.19 -1.65

sober 1 0.19 1 0 0.19 -1.65

soften 16 0.56 0.13 0.31 4.5 1.5

solidify 2 0.1 0.5 0.5 0.19 -1.65

sour 2 0.5 0.15 0.5 3.25 1.18

spin 7 0.29 0.43 0.29 0.67 -0.41

splay 0 0.19 0.31 0.5 0.63 -0.47

splinter 1 0.19 0.31 1 0.63 -0.47

split 117 0.24 0.05 0.71 4.67 1.54

sprout 5 0.2 0.8 0 0.25 -1.39

squeak 0 0.19 0.31 0.5 0.63 -0.47

squeal 0 0.19 0.31 0.5 0.63 -0.47

squirt 0 0.19 0.31 0.5 0.63 -0.47

stabilize 15 0.27 0.47 0.27 0.57 -0.56

stand 1349 0.15 0.85 0 0.17 -1.76

steady 2 0.1 0.5 0.5 0.19 -1.65

steep 15 0.07 0.02 0.93 3.25 1.18

steepen 0 0.19 0.31 0.5 0.63 -0.47

stiffen 1 1 0.31 0 3.25 1.18

stifle 69 0.57 0.04 0.39 13 2.56

straighten 9 0.33 0.22 0.44 1.5 0.41

stratify 0 0.19 0.31 0.5 0.63 -0.47

314

Verb Counts Caus.

rate

Anticaus.

rate

Passive

rate

C/A ra-

tio

Sp-value

strengthen 1670 0.52 0.05 0.43 10.47 2.35

stretch 90 0.41 0.24 0.34 1.68 0.52

submerge 12 0.02 0.08 0.92 0.19 -1.65

subside 10 0.02 1 0 0.02 -3.95

suffocate 25 0.32 0.28 0.4 1.14 0.13

sweeten 2 0.5 0.15 0.5 3.25 1.18

swim 20 0.01 1 0 0.01 -4.64

swing 15 0.13 0.87 0 0.15 -1.87

tame 4 0.75 0.08 0.25 9.75 2.28

tan 0 0.19 0.31 0.5 0.63 -0.47

taper 0 0.19 0.31 0.5 0.63 -0.47

tauten 0 0.19 0.31 0.5 0.63 -0.47

tear 95 0.31 0.03 0.66 9.67 2.27

tense 0 0.19 0.31 0.5 0.63 -0.47

thaw 0 0.19 0.31 0.5 0.63 -0.47

thicken 1 0.19 0.31 1 0.63 -0.47

thin 1 0.19 0.31 1 0.63 -0.47

thrill 11 0.18 0.03 0.82 6.5 1.87

tighten 175 0.38 0.07 0.55 5.58 1.72

tilt 7 0.43 0.29 0.29 1.5 0.41

tinkle 0 0.19 0.31 0.5 0.63 -0.47

tire 36 0.06 0.47 0.47 0.12 -2.14

topple 19 0.16 0.11 0.74 1.5 0.41

toughen 7 0.57 0.14 0.29 4 1.39

triple 19 0.21 0.74 0.05 0.29 -1.25

trot 15 0.2 0.13 0.67 1.5 0.41

turn 2003 0.37 0.48 0.15 0.77 -0.26

twang 0 0.19 0.31 0.5 0.63 -0.47

twirl 0 0.19 0.31 0.5 0.63 -0.47

twist 8 0.63 0.04 0.38 16.25 2.79

ulcerate 0 0.19 0.31 0.5 0.63 -0.47

315


Verb Counts Caus.

rate

Anticaus.

rate

Passive

rate

C/A ra-

tio

Sp-value

unfold 63 0.05 0.95 0 0.05 -3

unionize 0 0.19 0.31 0.5 0.63 -0.47

vaporize 0 0.19 0.31 0.5 0.63 -0.47

vary 159 0.11 0.84 0.05 0.14 -2

vibrate 0 0.19 0.31 0.5 0.63 -0.47

vitrify 2 0.1 0.5 0.5 0.19 -1.65

volatilize 0 0.19 0.31 0.5 0.63 -0.47

waken 1 0.19 1 0 0.19 -1.65

walk 76 0.16 0.84 0 0.19 -1.67

warm 10 0.1 0.9 0 0.11 -2.2

warp 5 0.6 0.06 0.4 9.75 2.28

weaken 435 0.53 0.09 0.38 5.75 1.75

weary 4 0.75 0.08 0.25 9.75 2.28

westernize 0 0.19 0.31 0.5 0.63 -0.47

whirl 1 0.19 0.31 1 0.63 -0.47

whiten 0 0.19 0.31 0.5 0.63 -0.47

widen 164 0.38 0.4 0.23 0.95 -0.05

wind 43 0.26 0.21 0.53 1.22 0.2

worry 597 0.29 0.46 0.26 0.63 -0.47

worsen 172 0.31 0.65 0.05 0.48 -0.74

wrinkle 0 0.19 0.31 0.5 0.63 -0.47

yellow 0 0.19 0.31 0.5 0.63 -0.47

ONE 26 0.19 0.31 0.5 0.63 -0.47

316

C. Verb aspect and event duration

data

317

C. Verb aspect and event duration data

Verb Pref. Suff. Asp. Dur.

believe 0.3 0.9 0.1 LONG

is 0.6 0.1 0.3 LONG

sold 0.9 0.3 0.7 LONG

deal 0.8 0.5 0.8 LONG

find 0.9 0.5 0.9 LONG

owns 0.8 0.8 0.2 LONG

crashed 0.2 0.6 0.6 LONG

thought 0.6 0.0 0.6 LONG

hit 0.7 0.1 0.7 LONG


spent 0.8 0.2 0.3 LONG

think 0.4 0.1 0.3 LONG

going 0.4 0.4 0.4 LONG


talking 0.9 0.1 0.3 LONG

estimates 0.7 0.3 0.7 LONG

is 0.6 0.1 0.3 LONG

going 0.4 0.4 0.4 LONG


lost 0.8 0.1 0.8 LONG

are 0.6 0.2 0.3 LONG

helping 0.7 0.3 0.3 LONG

fallen 0.4 0.1 0.9 LONG

are 0.6 0.2 0.3 LONG

turning 0.4 0.2 0.2 LONG

plunged 0.3 0.3 0.3 LONG

soared 0.7 0.3 0.3 LONG

said 0.2 0.8 0.9 SHORT

double 0.7 0.3 0.7 LONG

means 0.4 0.4 0.1 LONG

spending 0.9 0.1 0.4 LONG


get 0.7 0.4 0.8 LONG

goes 0.5 0.3 0.2 LONG

saw 0.2 0.1 0.8 SHORT

calls 0.4 0.2 0.2 SHORT

saw 0.2 0.1 0.8 SHORT

blew 0.8 0.2 0.8 SHORT

became 0.9 0.1 0.8 SHORT

released 0.2 0.2 0.6 SHORT

exploded 0.3 0.7 0.7 SHORT

went 0.7 0.4 0.8 SHORT


hear 0.1 0.9 0.9 SHORT

went 0.7 0.4 0.8 SHORT

tries 0.8 0.3 0.4 SHORT

asks 0.4 0.1 0.3 SHORT


make 0.6 0.1 0.7 LONG

give 0.6 0.3 0.6 LONG

kept 0.6 0.2 0.4 LONG

see 0.1 0.0 0.9 LONG

see 0.1 0.0 0.9 LONG

want 0.4 0.1 0.2 LONG

says 0.1 0.3 0.8 SHORT

invited 0.7 0.7 0.3 LONG

predicted 0.7 0.7 0.3 LONG

tried 0.9 0.3 0.5 LONG

said 0.2 0.8 0.9 LONG


endures 0.9 0.1 0.7 LONG

persuade 0.7 0.5 0.5 LONG

flew 0.7 0.3 0.7 LONG

318


say 0.2 0.3 0.9 SHORT

become 1.0 0.1 0.9 LONG

held 0.4 0.3 0.3 LONG

included 0.8 0.2 0.5 LONG

chosen 0.7 0.7 0.3 LONG

learned 0.9 0.1 0.8 SHORT

named 0.5 0.5 0.5 SHORT

taken 0.7 0.2 0.8 SHORT

hurried 0.6 0.2 0.6 LONG

makes 0.7 0.3 0.2 LONG

picked 0.8 0.6 0.9 LONG

followed 0.6 0.4 0.4 LONG

doing 0.3 0.2 0.4 LONG

was 0.1 0.0 0.1 LONG

has 0.6 0.2 0.7 LONG




saw 0.2 0.1 0.8 LONG

come 0.8 0.5 0.8 LONG


committed 0.9 0.1 0.9 LONG

expected 0.9 0.9 0.1 LONG


indicated 0.6 0.2 0.6 SHORT

told 0.5 0.5 0.7 SHORT

hopes 0.3 0.7 0.7 LONG

comes 0.3 0.7 0.2 LONG

arrive 0.6 0.6 0.8 LONG




shipping 0.4 0.2 0.4 LONG


stopped 0.8 0.1 0.8 LONG


number 0.6 0.3 0.3 LONG


met 0.6 0.2 0.9 LONG

bring 0.9 0.3 0.8 LONG

led 0.4 0.1 0.4 LONG


called 0.3 0.1 0.2 SHORT

appears 0.8 0.2 0.7 SHORT

say 0.2 0.3 0.9 SHORT

engaged 0.4 0.2 0.2 LONG

want 0.4 0.1 0.2 LONG

retreated 0.7 0.7 0.7 LONG




fear 0.4 0.2 0.4 LONG


denied 0.8 0.6 0.6 LONG


continue 0.8 0.2 0.6 LONG

receive 0.8 0.2 0.2 LONG

pushed 0.5 0.7 0.7 LONG

plans 0.5 0.8 0.5 LONG

occurred 0.5 0.3 0.9 LONG

say 0.2 0.3 0.9 SHORT

streamed 0.2 0.2 0.2 LONG

quoted 0.3 0.7 0.3 LONG

319



took 0.7 0.3 0.9 LONG

produce 0.7 0.2 0.5 LONG


continued 0.7 0.2 0.6 LONG









say 0.2 0.3 0.9 SHORT

declined 0.3 0.3 0.3 SHORT

included 0.8 0.2 0.5 LONG


grew 0.6 0.4 0.4 LONG

allows 0.3 0.3 0.7 LONG

seeking 0.2 0.2 0.2 LONG


ordered 0.7 0.3 0.7 LONG

talking 0.9 0.1 0.3 SHORT



has 0.6 0.2 0.7 LONG


close 0.8 0.1 0.4 LONG



finished 0.7 0.4 0.7 SHORT

is 0.6 0.1 0.3 LONG


fall 0.6 0.2 0.6 SHORT

has 0.6 0.2 0.7 LONG

hope 0.2 0.2 0.4 LONG


seen 0.1 0.1 0.9 LONG

want 0.4 0.1 0.2 LONG


need 0.6 0.3 0.4 LONG

understands 0.4 0.1 0.7 LONG

indicated 0.6 0.2 0.6 LONG

put 0.5 0.2 0.8 SHORT


following 0.5 0.3 0.5 LONG


save 0.2 0.2 0.8 LONG

save 0.2 0.2 0.8 LONG

begun 0.5 0.2 0.5 LONG

say 0.2 0.3 0.9 SHORT


hope 0.2 0.2 0.4 LONG

welcomed 0.7 0.3 0.7 LONG

is 0.6 0.1 0.3 LONG


becoming 0.8 0.2 0.5 LONG



taking 0.2 0.4 0.2 LONG

rests 0.8 0.2 0.6 LONG

shown 0.8 0.5 0.5 SHORT

diminishes 0.7 0.3 0.7 LONG

recognizes 0.8 0.1 0.9 LONG

320



announced 0.8 0.2 0.8 SHORT

leave 0.8 0.4 0.8 SHORT

add 0.8 0.2 0.5 LONG


provide 0.5 0.2 0.8 LONG


cut 0.8 0.6 0.8 LONG

double 0.7 0.3 0.7 LONG

appoint 0.7 0.3 0.7 LONG


working 0.3 0.3 0.2 LONG


working 0.3 0.3 0.2 LONG

led 0.4 0.1 0.4 LONG

began 0.2 0.0 0.8 LONG

captured 0.8 0.2 0.8 LONG


killed 0.8 0.1 0.6 SHORT


sent 0.5 0.2 0.7 LONG

turned 0.9 0.7 0.9 LONG

clearing 0.8 0.5 0.8 LONG

considered 0.4 0.2 0.2 LONG


found 0.9 0.6 0.9 LONG


buried 0.7 0.3 0.7 LONG

fled 0.8 0.5 0.8 LONG

falling 0.5 0.1 0.3 LONG

went 0.7 0.4 0.8 LONG


arrested 0.9 0.1 0.9 LONG



sent 0.5 0.2 0.7 LONG


received 0.7 0.1 0.3 LONG

appeared 0.7 0.2 0.7 LONG

find 0.9 0.5 0.9 LONG

involved 0.8 0.8 0.2 LONG

pushing 0.5 0.5 0.2 LONG



attacked 0.3 0.3 0.3 LONG



establish 0.7 0.3 0.7 LONG

happened 0.4 0.3 0.8 LONG

formed 0.7 0.7 0.4 LONG

died 0.9 0.1 0.7 LONG



believed 0.1 0.9 0.1 LONG

searching 0.7 0.3 0.7 LONG

entered 0.8 0.4 0.8 LONG

built 0.6 0.4 0.6 LONG

used 0.7 0.5 0.4 LONG

designed 0.8 0.2 0.8 LONG

gave 0.4 0.1 0.5 LONG

press 0.7 0.7 0.7 LONG


brought 0.8 0.1 0.9 LONG

321



rejected 0.7 0.3 0.7 SHORT

blocked 0.8 0.4 0.6 LONG

look 0.6 0.0 0.5 LONG


finish 0.8 0.2 0.8 LONG


touched 0.8 0.6 0.6 SHORT

shot 0.5 0.3 0.6 SHORT

agreed 0.4 0.2 0.9 LONG


consider 0.1 0.1 0.3 LONG

lived 0.1 0.2 0.1 LONG

require 0.2 0.2 0.2 LONG

remain 0.9 0.1 0.8 LONG

covered 0.9 0.4 0.4 LONG

discussed 0.3 0.3 0.7 LONG

quoted 0.3 0.7 0.3 SHORT

undermining 0.5 0.5 0.2 SHORT

claimed 0.1 0.1 0.1 LONG


permit 0.8 0.2 0.5 LONG


fell 0.5 0.1 0.7 LONG


signed 0.5 0.8 0.5 SHORT

ruled 0.5 0.2 0.2 LONG


claimed 0.1 0.1 0.1 LONG


say 0.2 0.3 0.9 SHORT

solved 0.2 0.5 0.5 LONG




assailed 0.7 0.3 0.7 LONG


create 0.2 0.2 0.5 LONG


join 0.8 0.2 0.8 LONG

marked 0.4 0.2 0.2 SHORT


had 0.5 0.2 0.6 LONG

had 0.5 0.2 0.6 LONG

secured 0.7 0.7 0.7 LONG

defend 0.3 0.3 0.3 LONG


gotten 0.7 0.3 0.7 LONG


crashed 0.2 0.6 0.6 SHORT

killed 0.8 0.1 0.6 LONG

made 0.8 0.2 0.7 SHORT

attacked 0.3 0.3 0.3 LONG


indicated 0.6 0.2 0.6 SHORT

cut 0.8 0.6 0.8 LONG

refer 0.8 0.5 0.5 SHORT



wants 0.3 0.2 0.3 LONG


claiming 0.1 0.1 0.1 LONG

sent 0.5 0.2 0.7 LONG

suffered 0.2 0.1 0.2 LONG

322



suggested 0.9 0.5 0.6 SHORT

occupies 0.8 0.2 0.2 LONG



calls 0.4 0.2 0.2 LONG




faced 0.7 0.1 0.6 LONG

created 0.5 0.2 0.8 LONG



lift 0.8 0.4 0.8 LONG

charged 0.2 0.2 0.5 LONG

stopped 0.8 0.1 0.8 LONG

inspected 0.8 0.2 0.2 LONG


resisted 0.9 0.3 0.4 LONG

appointed 0.7 0.3 0.7 LONG

mention 0.8 0.7 0.8 SHORT

agreed 0.4 0.2 0.9 SHORT

try 0.4 0.1 0.4 LONG



assisted 0.3 0.3 0.3 LONG

reported 0.5 0.8 0.5 SHORT

visited 0.3 0.3 0.3 LONG


finished 0.7 0.4 0.7 LONG

eliminated 0.2 0.8 0.8 LONG


seen 0.1 0.1 0.9 LONG


worked 0.3 0.1 0.2 LONG


led 0.4 0.1 0.4 LONG

blocked 0.8 0.4 0.6 LONG


arrived 0.7 0.5 0.8 LONG

destroyed 0.9 0.2 0.8 LONG


made 0.8 0.2 0.7 LONG

allowed 0.7 0.1 0.5 LONG

moved 0.9 0.6 0.6 LONG



argued 0.7 0.7 0.3 LONG

placed 0.6 0.2 0.4 LONG


visited 0.3 0.3 0.3 LONG

laid 0.8 0.1 0.8 SHORT


marched 0.6 0.4 0.4 LONG

say 0.2 0.3 0.9 SHORT

trying 0.8 0.5 0.4 LONG


chanted 0.5 0.5 0.5 SHORT


carried 0.5 0.1 0.5 LONG

found 0.9 0.6 0.9 LONG

appeared 0.7 0.2 0.7 LONG

prevented 0.8 0.2 0.8 LONG

323




pushed 0.5 0.7 0.7 LONG



marched 0.6 0.4 0.4 LONG

supposed 0.2 0.1 0.3 LONG


hit 0.7 0.1 0.7 SHORT

caused 0.9 0.6 0.9 SHORT

demanded 0.5 0.5 0.2 LONG


seized 0.9 0.1 0.9 SHORT

bombed 0.3 0.7 0.7 SHORT


kept 0.6 0.2 0.4 LONG

threatening 0.8 0.2 0.5 LONG

say 0.2 0.3 0.9 SHORT



want 0.4 0.1 0.2 LONG

move 0.5 0.5 0.5 LONG

sent 0.5 0.2 0.7 LONG

served 0.7 0.3 0.7 LONG

spent 0.8 0.2 0.3 LONG

expressed 0.7 0.7 0.3 SHORT

part 0.8 0.4 0.6 LONG

set 0.6 0.2 0.7 LONG



leaving 0.7 0.4 0.7 LONG



go 0.6 0.5 0.5 LONG

made 0.8 0.2 0.7 LONG

appointed 0.7 0.3 0.7 LONG


add 0.8 0.2 0.5 LONG

use 0.5 0.4 0.2 LONG

set 0.6 0.2 0.7 LONG

place 0.5 0.2 0.4 LONG

ensure 0.3 0.7 0.7 LONG

provide 0.5 0.2 0.8 LONG

emerge 0.5 0.2 0.8 LONG

emerged 0.8 0.2 0.8 LONG

help 0.8 0.4 0.8 LONG

has 0.6 0.2 0.7 LONG

beginning 0.7 0.3 0.3 LONG

come 0.8 0.5 0.8 LONG

reaching 0.5 0.5 0.8 LONG

create 0.2 0.2 0.5 LONG

prove 0.8 0.2 0.8 LONG

need 0.6 0.3 0.4 LONG

wrote 0.8 0.8 0.8 LONG

reported 0.5 0.8 0.5 LONG

organising 0.2 0.8 0.8 LONG

have 0.6 0.2 0.6 LONG


held 0.4 0.3 0.3 LONG

interpret 0.3 0.3 0.3 LONG

do 0.3 0.2 0.4 LONG

reproduced 0.3 0.7 0.7 SHORT

said 0.2 0.8 0.9 LONG


324



requires 0.3 0.3 0.3 LONG

wanted 0.3 0.2 0.3 LONG


gave 0.4 0.1 0.5 SHORT



cost 0.3 0.3 0.3 LONG


allow 0.8 0.2 0.6 LONG

took 0.7 0.3 0.9 LONG

approaching 0.8 0.6 0.4 LONG


waste 0.7 0.3 0.7 LONG



announced 0.8 0.2 0.8 LONG

been 0.6 0.2 0.6 LONG




hold 0.9 0.1 0.8 LONG

expect 0.6 0.4 0.6 LONG

authorized 0.7 0.3 0.7 LONG

stop 0.7 0.3 0.9 LONG


feels 0.8 0.1 0.3 LONG

happens 0.1 0.6 0.4 LONG

hope 0.2 0.2 0.4 LONG


have 0.6 0.2 0.6 LONG


say 0.2 0.3 0.9 SHORT

continues 0.7 0.1 0.5 LONG


committed 0.9 0.1 0.9 LONG

prepared 0.2 0.5 0.5 LONG

know 0.1 0.0 0.0 LONG


sought 0.3 0.3 0.3 SHORT

fails 0.6 0.1 0.9 LONG

presented 0.8 0.2 0.5 SHORT

arrived 0.7 0.5 0.8 LONG


fails 0.6 0.1 0.9 LONG

allow 0.8 0.2 0.6 LONG

wrapped 0.8 0.2 0.8 SHORT


added 0.9 0.1 0.9 SHORT

appears 0.8 0.2 0.7 LONG

come 0.8 0.5 0.8 LONG

fell 0.5 0.1 0.7 LONG

built 0.6 0.4 0.6 LONG

seeing 0.5 0.2 0.5 LONG

look 0.6 0.0 0.5 LONG

saw 0.2 0.1 0.8 LONG

made 0.8 0.2 0.7 LONG

falling 0.5 0.1 0.3 LONG

tells 0.3 0.5 0.7 LONG

discovered 0.7 0.1 0.9 LONG



hurt 0.5 0.2 0.5 SHORT

325



ignore 0.5 0.5 0.8 LONG

explodes 0.3 0.7 0.7 SHORT

say 0.2 0.3 0.9 SHORT

remember 0.2 0.0 0.3 LONG

claim 0.1 0.1 0.1 LONG

say 0.2 0.3 0.9 SHORT

suggesting 0.9 0.4 0.7 LONG

worked 0.3 0.1 0.2 LONG

charged 0.2 0.2 0.5 LONG

suspect 0.8 0.2 0.8 LONG


ordered 0.7 0.3 0.7 SHORT


came 0.8 0.4 0.7 LONG

left 0.9 0.1 0.6 LONG



hear 0.1 0.9 0.9 LONG

put 0.5 0.2 0.8 LONG

causing 0.7 0.7 0.7 LONG

returning 0.2 0.2 0.8 LONG

want 0.4 0.1 0.2 LONG

happens 0.1 0.6 0.4 LONG

appears 0.8 0.2 0.7 LONG

coming 0.9 0.6 0.4 LONG

come 0.8 0.5 0.8 SHORT



go 0.6 0.5 0.5 SHORT


are 0.6 0.2 0.3 LONG


are 0.6 0.2 0.3 LONG


lost 0.8 0.1 0.8 SHORT

found 0.9 0.6 0.9 SHORT

contain 0.8 0.2 0.5 LONG



have 0.6 0.2 0.6 LONG

live 0.2 0.1 0.3 LONG

declared 0.7 0.3 0.7 SHORT

flying 0.5 0.2 0.5 LONG

added 0.9 0.1 0.9 SHORT


reflected 0.9 0.2 0.3 LONG

helped 0.8 0.4 0.4 LONG

showed 0.7 0.5 0.3 LONG

added 0.9 0.1 0.9 LONG

remained 0.9 0.1 0.8 LONG

became 0.9 0.1 0.8 LONG

remained 0.9 0.1 0.8 LONG

fallen 0.4 0.1 0.9 LONG


reflected 0.9 0.2 0.3 LONG

creating 0.3 0.3 0.3 LONG

hit 0.7 0.1 0.7 LONG

reflects 0.9 0.2 0.4 LONG

added 0.9 0.1 0.9 LONG

came 0.8 0.4 0.7 LONG

led 0.4 0.1 0.4 LONG



326



move 0.5 0.5 0.5 LONG

broke 0.8 0.2 0.8 LONG


dropped 0.8 0.3 0.8 LONG


withstood 0.7 0.3 0.7 LONG


caused 0.9 0.6 0.9 LONG

extended 0.7 0.3 0.3 LONG

rose 0.4 0.2 0.5 LONG

suggested 0.9 0.5 0.6 LONG

intend 0.7 0.7 0.3 LONG

set 0.6 0.2 0.7 LONG


edged 0.7 0.3 0.3 LONG


rose 0.4 0.2 0.5 LONG

welcomed 0.7 0.3 0.7 LONG


expects 0.9 0.8 0.2 LONG

made 0.8 0.2 0.7 LONG

raise 0.8 0.7 0.8 LONG


disregarded 0.7 0.3 0.7 SHORT


running 0.5 0.2 0.3 LONG


left 0.9 0.1 0.6 LONG

extending 0.6 0.2 0.4 LONG

given 0.6 0.5 0.9 LONG


rose 0.4 0.2 0.5 LONG

lost 0.8 0.1 0.8 LONG

exhaust 0.7 0.7 0.3 LONG

start 0.5 0.4 0.6 LONG

declined 0.3 0.3 0.3 LONG



believes 0.2 0.9 0.1 LONG

asked 0.6 0.1 0.3 SHORT

want 0.4 0.1 0.2 SHORT

made 0.8 0.2 0.7 SHORT

know 0.1 0.0 0.0 LONG

stated 0.4 0.8 0.8 SHORT

pleased 0.3 0.1 0.1 SHORT

removed 0.6 0.2 0.6 LONG

intended 0.7 0.2 0.3 LONG

lives 0.8 0.3 0.8 LONG


took 0.7 0.3 0.9 SHORT

waiting 0.1 0.1 0.1 SHORT

looks 0.7 0.3 0.3 LONG




finished 0.7 0.4 0.7 SHORT

emptied 0.8 0.2 0.8 SHORT

appeared 0.7 0.2 0.7 SHORT


watching 0.9 0.1 0.1 SHORT


327



fled 0.8 0.5 0.8 SHORT


appear 0.9 0.3 0.7 LONG

started 0.2 0.4 0.8 SHORT

hit 0.7 0.1 0.7 SHORT

pronounced 0.7 0.3 0.3 SHORT

seen 0.1 0.1 0.9 LONG

related 0.7 0.3 0.3 LONG

expressed 0.7 0.7 0.3 LONG

believed 0.1 0.9 0.1 LONG



arrested 0.9 0.1 0.9 LONG



identified 0.7 0.3 0.7 LONG


made 0.8 0.2 0.7 LONG

needed 0.2 0.2 0.1 LONG


beaten 0.6 0.6 0.3 SHORT




seem 0.4 0.4 0.4 LONG

have 0.6 0.2 0.6 LONG

asked 0.6 0.1 0.3 SHORT

argued 0.7 0.7 0.3 SHORT

cover 0.7 0.3 0.7 LONG

retreated 0.7 0.7 0.7 LONG

supported 0.7 0.7 0.3 LONG


agree 0.8 0.2 0.3 LONG

took 0.7 0.3 0.9 LONG


expressed 0.7 0.7 0.3 SHORT

written 0.4 0.4 0.6 LONG

found 0.9 0.6 0.9 SHORT

chose 0.8 0.2 0.8 LONG

make 0.6 0.1 0.7 LONG


seized 0.9 0.1 0.9 LONG

released 0.2 0.2 0.6 LONG

thrown 0.8 0.2 0.8 LONG


bought 0.1 0.1 0.9 LONG

convicted 0.7 0.7 0.3 LONG

has 0.6 0.2 0.7 SHORT

facing 0.5 0.5 0.5 LONG

reinstated 0.3 0.3 0.7 LONG

raised 0.8 0.7 0.7 LONG


stand 0.4 0.2 0.5 LONG

speak 0.5 0.0 0.6 LONG

coming 0.9 0.6 0.4 SHORT

destroyed 0.9 0.2 0.8 LONG

kept 0.6 0.2 0.4 LONG


used 0.7 0.5 0.4 LONG

think 0.4 0.1 0.3 SHORT

got 0.6 0.2 0.7 SHORT

have 0.6 0.2 0.6 LONG

stand 0.4 0.2 0.5 LONG

328


rolled 0.6 0.2 0.4 SHORT

discussed 0.3 0.3 0.7 LONG


abandoned 0.8 0.2 0.8 LONG

took 0.7 0.3 0.9 SHORT

go 0.6 0.5 0.5 LONG

demonstrating 0.7 0.3 0.7 LONG

ordered 0.7 0.3 0.7 SHORT

hurt 0.5 0.2 0.5 SHORT

get 0.7 0.4 0.8 LONG

say 0.2 0.3 0.9 SHORT

slipping 0.4 0.6 0.8 LONG

say 0.2 0.3 0.9 SHORT

destroy 0.9 0.5 0.5 SHORT

cascaded 0.7 0.3 0.7 SHORT

break 0.4 0.1 0.6 SHORT


say 0.2 0.3 0.9 SHORT

warning 0.7 0.7 0.3 SHORT

cause 0.8 0.4 0.8 LONG

reports 0.3 0.7 0.7 SHORT

fall 0.6 0.2 0.6 SHORT


destroy 0.9 0.5 0.5 SHORT

used 0.7 0.5 0.4 LONG

presented 0.8 0.2 0.5 LONG





called 0.3 0.1 0.2 LONG


fight 0.1 0.1 0.1 LONG

changing 0.8 0.2 0.5 LONG

asked 0.6 0.1 0.3 LONG

bought 0.1 0.1 0.9 LONG

had 0.5 0.2 0.6 LONG

decide 0.8 0.2 0.8 LONG

move 0.5 0.5 0.5 LONG

laid 0.8 0.1 0.8 LONG

dropped 0.8 0.3 0.8 SHORT

delivered 0.3 0.3 0.7 SHORT

has 0.6 0.2 0.7 LONG



has 0.6 0.2 0.7 LONG



ignore 0.5 0.5 0.8 LONG

say 0.2 0.3 0.9 SHORT

rule 0.5 0.5 0.2 LONG

embrace 0.7 0.3 0.7 LONG

providing 0.4 0.2 0.6 LONG

bring 0.9 0.3 0.8 LONG

say 0.2 0.3 0.9 SHORT

led 0.4 0.1 0.4 LONG

teaches 0.8 0.2 0.4 LONG

learning 0.7 0.3 0.7 LONG



has 0.6 0.2 0.7 LONG

take 0.6 0.2 0.7 LONG

denounced 0.9 0.3 0.7 LONG

329



complained 0.7 0.7 0.7 SHORT

wondered 0.4 0.1 0.3 SHORT


shifted 0.7 0.7 0.7 LONG

fear 0.4 0.2 0.4 LONG

worried 0.3 0.7 0.3 LONG

saying 0.5 0.3 0.5 SHORT


taking 0.2 0.4 0.2 LONG

reports 0.3 0.7 0.7 SHORT



say 0.2 0.3 0.9 LONG

state 0.4 0.8 0.8 SHORT


reports 0.3 0.7 0.7 LONG


says 0.1 0.3 0.8 LONG





seen 0.1 0.1 0.9 LONG

received 0.7 0.1 0.3 SHORT

using 0.4 0.4 0.5 LONG

jumped 0.7 0.3 0.7 LONG

say 0.2 0.3 0.9 SHORT

approaching 0.8 0.6 0.4 LONG


have 0.6 0.2 0.6 LONG

becoming 0.8 0.2 0.5 LONG


beginning 0.7 0.3 0.3 LONG

used 0.7 0.5 0.4 LONG

rise 0.5 0.7 0.7 LONG


identified 0.7 0.3 0.7 LONG

warns 0.7 0.7 0.3 SHORT

grew 0.6 0.4 0.4 LONG




attracted 0.7 0.7 0.7 LONG

opening 0.6 0.2 0.2 LONG

surprising 0.7 0.3 0.7 LONG


joining 0.8 0.5 0.8 LONG




puts 0.6 0.2 0.7 LONG






appealing 0.7 0.3 0.3 LONG



make 0.6 0.1 0.7 LONG


organized 0.2 0.8 0.8 LONG


330




facing 0.5 0.5 0.5 LONG

forbidden 0.8 0.2 0.8 LONG

have 0.6 0.2 0.6 LONG

torn 0.4 0.4 0.4 LONG

gets 0.3 0.7 0.3 LONG

forced 0.8 0.5 0.3 LONG


runs 0.5 0.2 0.6 LONG


put 0.5 0.2 0.8 LONG

do 0.3 0.2 0.4 LONG

works 0.3 0.7 0.3 LONG

gave 0.4 0.1 0.5 LONG



served 0.7 0.3 0.7 LONG

made 0.8 0.2 0.7 LONG

works 0.3 0.7 0.3 LONG



need 0.6 0.3 0.4 LONG

have 0.6 0.2 0.6 LONG

treated 0.8 0.2 0.5 LONG



waiting 0.1 0.1 0.1 LONG

do 0.3 0.2 0.4 LONG

save 0.2 0.2 0.8 LONG

begins 0.8 0.2 0.2 LONG


is 0.6 0.1 0.3 LONG

accumulate 0.7 0.3 0.7 LONG

see 0.1 0.0 0.9 LONG


asked 0.6 0.1 0.3 LONG

expects 0.9 0.8 0.2 LONG

suggested 0.9 0.5 0.6 SHORT



meet 0.5 0.2 0.7 LONG


paid 0.9 0.4 0.2 LONG

had 0.5 0.2 0.6 LONG

purchased 0.3 0.3 0.7 LONG


denied 0.8 0.6 0.6 LONG


decided 0.5 0.2 0.8 SHORT

turned 0.9 0.7 0.9 SHORT

came 0.8 0.4 0.7 SHORT


produced 0.9 0.1 0.5 LONG

approved 0.7 0.3 0.7 SHORT


begun 0.5 0.2 0.5 LONG


narrowed 0.7 0.7 0.3 LONG




closed 0.7 0.2 0.5 SHORT

331



issue 0.8 0.2 0.8 LONG


adopted 0.8 0.2 0.8 LONG

had 0.5 0.2 0.6 LONG

have 0.6 0.2 0.6 LONG

was 0.1 0.0 0.1 LONG

suffered 0.2 0.1 0.2 LONG







made 0.8 0.2 0.7 LONG

help 0.8 0.4 0.8 LONG

profit 0.7 0.3 0.3 LONG


rose 0.4 0.2 0.5 LONG





include 0.8 0.2 0.4 LONG

close 0.8 0.1 0.4 LONG


result 0.8 0.5 0.2 LONG


rose 0.4 0.2 0.5 LONG



fell 0.5 0.1 0.7 LONG


has 0.6 0.2 0.7 LONG


closed 0.7 0.2 0.5 SHORT

improve 0.7 0.3 0.3 LONG


apply 0.7 0.3 0.2 LONG



issue 0.8 0.2 0.8 LONG

exercised 0.8 0.2 0.5 SHORT



has 0.6 0.2 0.7 LONG

issued 0.8 0.3 0.7 LONG



paid 0.9 0.4 0.2 LONG

had 0.5 0.2 0.6 LONG

purchased 0.3 0.3 0.7 LONG


332

dynamics, causation, duration in the predicate-argument ... · universit e de gen eve facult e des...

Documents