Bruckenanaphern und die Integration vonSDRT mit FrameNet

Matthias Irmermirmer@uni-leipzig.de

Semantik-Kolloquium, Institut fur Linguistik, Universitat Leipzig

27. Mai und 3. Juni 2010

Contents

1 Bridging Anaphora 21.1 A Preliminary Classification of Bridging Anaphora . . . . . . . . . . . . . . . . . 21.2 Corpus Studies on Anaphoric Expressions . . . . . . . . . . . . . . . . . . . . . . 31.3 Is Bridging A Lexical Phenomenon? . . . . . . . . . . . . . . . . . . . . . . . . . 41.4 A Refined Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 Bridging in SDRT 72.1 The Discourse Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72.2 Representing Bridging Anaphora . . . . . . . . . . . . . . . . . . . . . . . . . . . 72.3 Bridging as Byproduct of Computing Discourse Structure . . . . . . . . . . . . . 92.4 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92.5 Another Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

3 Frame Semantics and FrameNet 133.1 Basic Units . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133.2 Frame Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133.3 Relations between Frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153.4 Relations between Frame Elements . . . . . . . . . . . . . . . . . . . . . . . . . . 17

4 Marrying SDRT with FrameNet 174.1 Proposal: Integrate FrameNet and SDRT . . . . . . . . . . . . . . . . . . . . . . 174.2 Representing Frame Elements in SDRT . . . . . . . . . . . . . . . . . . . . . . . 184.3 Establishing Discourse Relations by FrameNet Data . . . . . . . . . . . . . . . . 20

4.3.1 Inferring Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204.3.2 Inferring Narration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

5 Constraints on Bridging Inferences 215.1 Problems for Resolving Bridging Anaphora . . . . . . . . . . . . . . . . . . . . . 215.2 Bridging Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

5.2.1 The Preference for Coreference . . . . . . . . . . . . . . . . . . . . . . . . 215.2.2 The Plausibility Constraint . . . . . . . . . . . . . . . . . . . . . . . . . . 225.2.3 The Right Frontier Constraint . . . . . . . . . . . . . . . . . . . . . . . . 225.2.4 Maximize Discourse Coherence . . . . . . . . . . . . . . . . . . . . . . . . 23

5.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

1

Introduction

• natural language discourses are structured: relations between utterances at various levels

• basic distinction between coherence and cohesion

coherence: rhetorical relations between text segments

cohesion: anaphoric relations between discourse entitites

1 Bridging Anaphora

• in a bridging anaphor, an entity introduced in a discourse stands in a particular relationto some previously mentioned discourse entity

• the Bridging Relation (Clark, 1977)

– is not explicitly stated– is an essential part of the discourse content: the knowledge of this relation is needed

for successfully interpreting a discourse

1.1 A Preliminary Classification of Bridging Anaphora

Clark (1977) distinguishes between direct reference and indirect reference

direct reference : the antecedent is an entity just mentioned, as in (1).

(1) I met a man yesterday. He told me a story.

indirect references by association : the antecedent is not directly mentioned, but closely asso-ciated with an entity mentioned beforethe associated pieces of information vary in their predictability

• necessary parts(2) I looked into the room. The ceiling was very high.

• probable parts(3) I walked into the room. The windows looked out to the bay.

• inducible parts(4) I walked into the room. The chandeliers sparkled brightly.

indirect reference by characterization : the bridging relation characterizes a role that some-thing implicitly plays in an event or circumstance mentioned beforeroles can be agents, objects, or instruments

• necessary roles(5) John was murdered yesterday. The murderer got away.

• optional roles(6) John was murdered yesterday. The knife lay nearby.

reasons, causes, consequences, concurrences and other relations between events

• not really bridging but rather a phenomenon of discourse coherence (rhetorical rela-tions)• Clark’s example for consequences is (7)

(7) John fell. What he did was break his arm.

• the two utterances are connected by a coherence relation Result

2

1.2 Corpus Studies on Anaphoric Expressions

• bridging inferences can be trigerred by

definite noun phrases

indefinite noun phrases

Definite Noun Phrases

• most classifications of bridging anaphora start from the observation that bridging relationsare triggered mainly by definite descriptions

• Russell (1905): definite descriptions are characterized by two properties: uniqueness andexistence.

Uniqueness (Russell, 1905)

• the referent denoted by the definite description must be the only referent that satisifiesthe given description• always relative to some restricted context, e.g. a quantification domain or the dis-

course domain

Familiarity • the referent denoted by the definite description must be known to the hearer• subsumes Russell’s condition of the existence of a referent• Prince (1981): familiarity hierarchy

discourse-old hearer-olddiscourse-new hearer-new

Indefinite Noun Phrases

• there are also cases of referring indefinite noun phrases which convey a bridging relation

(8) John was murdered yesterday. A knife lay nearby.

• “a knife” clearly refers to the probable killing instrument, almost identically as in (6)

• uniqueness is not a necessary condition for bridging, but some sort of familiarity is involved

Empirical investigations on the use of definite descriptions

• Fraurud (1990): classification of definite descriptions based on Swedish corpus data

indefinite NPs definite NPsfirst-mentioned 85.2% 60.9%subsequent-mentioned 8.3% 36.1%others 6.5% 3.0%

• Poesio and Vieira (1998): classification based on native-speaker annotations (in a pre-defined classification scheme) of a corpus of English newspaper articles

3

Class Poesio & Vieira Frequency Prince Fraurud RobertsI coreferential NPs 43-45% discourse old subsequent-

mentioned(39%)

stronglyfamiliar

II bridging 6-11% discourse new, but re-lated to a discourse-oldentity

first-mentioned(61%)

weaklyfamiliar

III larger situation 20-25% discourse new andhearer-old

IV unfamiliar 18-26% discourse-new andhearer-new andnot related to anydiscourse-old entity

• Gardent et al. (2003): corpus investigation and subclassification of bridging anaphora

– annotation of definite descriptions

Relation Number PercentageCoreferential 1481 16.85%Bridging 416 4.73%First mention 6892 78.40%

– annotation of bridging definite descriptions

Class Occurrences Sourceset membership 5.8% hyponymythematic 5.3% event thematic griddefinitional 80.8% lexicographic definitionco-participants 2.2% lexicographic definitionsnon-lexical 7.8% discourse structure, world knowledge

– problem: availability of a well-sorted information source that could provide lexico-graphic definitions needed in the definitional and co-participants classes

– suggestions- meronymic relations (52%) could be processed using WordNet, but only 38 of

187 meronymic cases were present in WordNet- thematic relations (5.3%) could be processed by FrameNet, in 14 of 19 the-

matic cases, a frame containing a target and an anchor as frame element were found

1.3 Is Bridging A Lexical Phenomenon?

• Schwarz (2000); Schwarz-Friesel (2007): a more recent classification of bridging (or indi-rect) anaphora with a cognitive emphasis

• four basic types of indirect anaphora are distinguished

1. Mereological semantic indirect anaphora involve a form of part-whole relationshipsbetween anchor and anaphor

- these relations are part of common semantic knowledge

(9) Er bemerkte jedoch im gleichen Augenblick, dass der Mann tot war. Die Schlafenwaren durchschossen.

’At that particular moment he noticed that the man was dead. His temples were shotthrough.’ ([from a novel of Friedrich Durrenmatt (1950)]) (Schwarz-Friesel, 2007, p. 9)

4

2. Lexical/thematic semantic indirect anaphora are based on lexical knowledge or onthematic roles

(10) Ich wollte rasch die Haustur aufschließen, weil ich das Telefon klingeln horte. DerSchlussel war aber tief unten im Einkaufswagen vergraben.

’I wanted to unlock the door quickly, because I could hear the telephone ringing. Thekey, however, was buried deeply in the trolley.’ (Schwarz-Friesel, 2007, p. 9)

3. Frame-/script-based conceptual indirect anaphora base their interpretation on theactivation of frame or script knowledge

(11) Ich kenne ein schones Restaurant in Refrath. Das Essen ist kostlich, und der Kellnerganz besonders nett.

’I know a lovely restaurant in Refrath. The food is excellent and the waiter is anextremely nice guy.’ (Schwarz-Friesel, 2007, p. 9)

4. Inference-based conceptual indirect anaphora require complex inferencing

(12) One night a man rushes into the police station and tells the policemen that he has justbeen knocked down in his garden. One policeman is asked to go and look for traces atthe place of the assault. After a short time he returns with a huge swelling at his headand says “I solved the case.” – “Bravo,” says his boss, “and how did you do that?” –“I stepped on the rake, too.” (Schwarz-Friesel, 2007, p. 10)

• A Problem: How to tell apart lexical from frame-based bridging, especially for cases wherethe anchor is a verb?

• the distinction of semantic vs. conceptual (i.e. pragmatic) bridging anaphora is not thatclear-cut because defeasible (i.e. pragmatic) inferences may be necessary in all types ofindirect anaphora, including mereological bridging anaphora

• a possible view: the bridging relation is part of the lexical entry of the anchor (e.g. Boset al. (1995))

• question: Is the bridging relation part of the lexical knowledge associated with bridginganchors, especially when the anchor is a verb?

• Garrod and Terras (2000) carried out an eye-tracking experiment in order to addressthe question whether the bridging relation between verb and anaphor (role-filler)

- is mediated by the context in which the verb occurs (“contextual account”)

- or rather results from a purely lexical association between the verb and its role-filler(“lexical account”)

• they compared the processing of target sentences (14) when following one of the sentencesin (13)

(13) a. The teacher was busy writing a letter of complaint [with a pen / to a parent].

b. The teacher was busy writing an exercise on the blackboard [with chalk / ∅].

(14) However, she was disturbed by a loud scream from the back of the class and the[pen/chalk] dropped on the floor.

• two main findings

5

1. first-pass reading times of the target sentence are slower when following an implicitintroduction of the thematic argument only for nondominant verb-role pairs (e.g.“chalk” in the context of “write”); no difference was found between explicit andimplicit antecedents for dominant verb-role pairs (“pen” in the context of “write”)

2. an early context effect for the dominant but not the nondominant verb-role pairs:reading times were faster following appropriate contexts (“write a letter” for “pen”)than following inappropriate contexts (“write on the blackboard” for “pen”); no sucheffect was observed with nondominant verb-role pairs. However, a strong contexteffect emerged later, in the second-pass reading times on the noun, both for dominantand nondominant targets

• Garrod and Terras (2000) suggest that their results indicate a two-stage process of dis-course role resolution (cf. also Sanford et al. 1993)

1. bonding stage- a link between the referring expression and a previous verb is established- a low-level automatic process not influenced by the particular context in which

the verb occurred2. resolution stage

- a process which evaluates the link made in the bonding process with respectto the overall discourse context, recomputes it if necessary, and integrates it into thediscourse model

• although the findings regarding the first-pass reading times could be interpreted as favoringa purely lexical approach, they can be explained simply by the fact that the collocationpattern “write” and “pen” is more frequent than the collocation of “write” and “chalk”

• lexical view cannot account for the second process involving broad contextual knowledge

• conclusion: bridging relations are mediated by the context in which the verb occurs andthus not likely to be stored in the lexicon

• suggestion: in the early bonding process underspecified material is filled by default values,which undergo a consistency check in the later resolution process

1.4 A Refined Classification

Mereological Bridging Anaphora

• the anchor is an already established discourse referent, its type is either an entity or a setof entities

• there is a part-of relation between anaphor and anchor

• e.g. part-of, element-of or any other mereological relation (member/collection, subcollec-tion/collection, portion/whole, substance/whole, component/integral whole)

(15) We went by car to Leipzig. The motor broke down.

Frame-related Bridging Anaphora

• the anchor is an eventuality or a frame present in the discourse model

• the bridging relation can be

6

1. either a thematic relation (e.g. agent-of, instrument-of, theme-of ) which the entity denotedby the anaphor plays in the eventuality

2. or another conceptual relation (causal, temporal, spatial)

(16) I was driving to Leipzig. The car broke down.

Bridging class Anchor type Typical bridging relationsmereological (sets of) entities part-of, member-offrame-related eventualities thematic: agent, instrument, theme

causal: cause-of, spatial: in, temporal...

2 Bridging in SDRT

2.1 The Discourse Model

• in a successful interpretation, all information, not only directly expressed but also indi-rectly inferred, will be part of the discourse model constructed by the hearer in course ofinterpretation

• interpretation involves constructing incrementally a structured mental representation ofthe discourse

• Cornish, 1999: the discourse model is “a constantly evolving representation of the entities,propositions, eventualities, properties, and states, as well as their interrelations, which areintroduced into the discourse, or are assumed already to exist therein, at particular points”

• most theories of discourse structure make use of a notion of discourse relations (orrhetorical relations or coherence relations)

Segmented Discourse Representation Theory (Asher and Lascarides, 2003)

• roughly: SDRT = dynamic semantics (DRT, Kamp and Reyle, 1993) + two new expres-sions

1. speech act discourse referents which label content of text segments and keeptrack of token utterances

2. rhetorical relations which relate speech act discourse referents

• the resulting structures are segmented DRSs (SDRSs)

2.2 Representing Bridging Anaphora

• Russell (1905): the denotation of a definite noun phrase can only be given if it fulfills theconditions of existence and uniqueness

• the condition of existence is usually expressed by ∃x[P (x)]

• the condition of uniqueness by ∃x∀y[P (y)→ x = y]

• the ι operator maps a set containing only one element to this element:

(17) Q(ιx[P (x)]) = ∃x[P (x) ∧ ∀y[P (y)→ y = x] ∧Q(x)]

• semantic representation of the definite article “the”:

(18) λPλQ[Q(ιx[P (x)])]

7

• Chierchia (1995, p. 221) extends this notion and includes a contextual parameter B for abridging relation

• “the P” denotes a P that is related by B to an antecedent a to be specified by context

• B restricts the domain and must be included in the uniqueness condition

• refined semantic representation of the definite article “the”:

(19) λPλQ[Q(ιx[B(a, x) ∧ P (x)])]

• the meaning of a definite noun phrase “the P” is characterized by

(20) a. λQ[Q(ιx[B(a, x) ∧ P (x)])] or

b. λQ[∃x[B(a, x) ∧ P (x) ∧ ∀y[P (y)→ y = x] ∧Q(x)]]

• building on that, Asher and Lascarides (1998a, p. 87) give a SDRT representation for adefinite noun phrase corresponding to (20) as shown in (21).

(21) λQ

π, v,R

π :

x, a,B

Q(x), P (x),B(a, x), B =?, a =?

y

P (y)B(a, y)

⇒y = x

R(v, π), R =?, v =?

• the predicates P (x) and Q(x) representing the meanings of the noun and the verb aretranslated as conditions on a discourse referent x

• the bridging relation B(a, x) is represented as an underspecified predicate in the set ofSDRS-conditions; the bridging anchor a shows up as an underspecified discourse referent

• the condition of uniqueness is now represented by the complex SDRS-condition consistingof the two small SDRSs connected by ⇒

• shortcut: for a discourse referent x introduced by a definite NP, just write !x in theuniverse, and omit the complex SDRS-condition representing the uniquess condition

• the representation of an indefinite noun phrase triggering a bridging inference would bevery similar: just leave out the uniqueness condition and keep the rest of the conditions

• the usual semantic representation for the indefinite article “a” is

(22) λPλQ∃x[P (x) ∧Q(x)]

• extending this formula by a bridging relation, we get

(23) λPλQ∃x[B(a, x) ∧ P (x) ∧Q(x)]

• two underspecifications are to be specified by pragmatic inference:

8

1. the coherence relation R(u, v) (Asher and Lascarides, 1998b)- a definite description triggers a coherence relation between the current utterance

u and some previous utterance v2. the bridging relation B(a, x) (Asher and Lascarides, 1998a)

- the parameters B and a have to be specified- for direct anaphora, B is identity- for mereological bridging, B can be part-of or member-of- for frame-related bridging, B is a thematic role, e.g. agent, theme, or instrument,

or a conceptual relation, e.g. cause, location, etc.

2.3 Bridging as Byproduct of Computing Discourse Structure

• in SDRT, bridging inferences are seen as “a byproduct of computing how the currentsentence connects to the previous ones in the discourse” (Asher and Lascarides, 1998a)

• four meta-rules for bridging:

1. If possible use identity.– reflects the empirical preference of resolving anaphora to an identical antecedent– preferred rule; if resolution to identity is not possible, then the other rules apply

in the indicated order2. Bridges must be plausible.

– world knowledge “specifies certain plausible ways of filling the underspecifiedparameters in the presupposed material”

– plausibility relies on world knowledge, but is not defined precisely3. Discourse structure determines bridging.

– if a rhetorical relation between the discourse segments involved gives particularclues for resolving the anaphora, then this information is to be used

4. Maximize discourse coherence.– one of the most basic principles assumed in SDRT– in discourse interpretation, there is a preference for resolving bridging anaphora

in a way that maximizes discourse coherence

2.4 An Example

(24) a. Peter zog von Leipzig nach Munchen.

b. Die Miete war billiger.

(25) a.

π1

π1 :

p, l,m, e1

named(p, peter)leipzig(l),munich(m)e1 : move(p, l,m)

9

b.

π2, R, v

π2 :

r, e2, B, a

rent(r)e2 : low(r)B(a, r)B =?, a =?

r′

rent(r′)B(a, r′)

⇒r′ = r

R(v, π2), R =?, v =?

• the first meta-rule for bridging – resolve to identity – cannot be applied because there isno available antecedent that could be coreferential to the referent r

• for the second rule – plausibility – we have to consider world knowledge

• on the one hand, it is well known in Germany that rents are, at least at the moment, muchless expensive in Leipzig than in Munich

(26) a. (rent(r) ∧ leipzig(l) ∧ in(l, r)) > low(r)b. (rent(r) ∧munich(m) ∧ in(m, r)) > ¬low(r)

• on the other hand, it is common world knowledge that a lower rent is a cause to move.We can express this by another default rule:

(27) (rent(r) ∧ in(r, b) ∧ eβ : low(r) ∧ eα : move(x, a, b)) > causeD(β, α)

• world knowledge gives rise to conflicting interpretations!

• the third meta-rule – discourse structure determines bridging: there are (at least) twodiscourse relations which are plausible to hold

Background (?(α, β, λ) ∧ event(eα) ∧ state(eβ)) > Background(α, β, λ)

Explanation (?(α, β, λ) ∧ causeD(β, α)) > Explanation(α, β, λ)1.

• a way out of this conflict is to impose an ordering �τ on coherence relations based on their“quality”

(28) Explanation(α, β, λ) �τ Background(α, β, λ)

• the last meta-rule – maximize discourse coherence (MDC) – tells us that “better” relationsare to be preferred

• we can specify R as Explanation and resolve B to in and a to m

• a pragmatically enriched SDRS for text (24) in its preferred reading:

1causeD(β, α) means that there is evidence in the discourse that β caused α.

10

(29)

π1, π2

π1 :

p, l,m, e1

named(p, peter)leipzig(l),munich(m)e1 : move(p, l,m)

π2 :

r, e2

rent(r)e2 : low(r)in(m, r)

r′

rent(r′)in(m, r′)

⇒r′ = r

Explanation(π1, π2)

2.5 Another Example

(30) John entered a room. The chandelier sparkled brightly.

(31) a.

π1

π1 :

e1, j, r

e1 : enter(j, r)named(j, john)room(r)

b.

π2, v, R

π2 :

e2, !x, a,B

e2 : sparkle(x)chandelier(x)bright(e2)B(x, a)a =?, B =?

R(v, π2), R =?, v =?

• world knowledge: rooms normally have a source of light

(32) room(r) > light(l) ∧ part− of(l, r)

• world knowledge: chandeliers are sources of light

(33) chandelier(x) > light(x)

• application of (32) enriches the SDRS of the first sentence by a referent l and the conditionslight(l) and part− of(l, r)

• rule (33) introduces a condition light(x) in the second SDRS

• the meta-rules for bridging, as they are stated, are not sufficient to resolve the bridgingrelation, because a link between x and l has to be assumed, and neither the discoursestructure nor the MDC can provide clues for that

• a solution: extending the MDC by a minimality criterion

• can be achieved by a general, low-ranked default “Equality by Default” (Cohen, 2007):

(34) > x = y

• a first attempt to adapt it to SDRT:

11

Equality by Default 2

πi : x ∧ πj : y ∧ (x is available in πj) > x = y

• this default provides a way to achieve minimality of discourse representations

• “minimize labels” permits to assume the equivalences x = l and a = r

• “minimize DRS-conditions” leads to B = part− of

• in order to connect the second SDRS by a discourse relation R to the discourse, look atthe eventualities in the text

• “to enter a room” is an achievement (in the terminology of Vendler, 1957)

• “to sparkle” is, in a strict sense, an activity; however, as it is a durative instance of emittinglight, it can be seen as a special case of a state (in a broad sense)

• assuming this, we are able to apply SDRT’s default rule for inferring a Backgroundrelation (cf. page 10 above)

• pragmatically enriched SDRS for (30):

(35)

π1, π2

π1 :

e1, j, r, l

e1 : enter(j, r)named(j, john)room(r)light(l)part− of(l, r)

π2 :

e2, !x, a,B

e2 : sparkle(x)chandelier(x), light(x)bright(e2)B(x, a)x = la = r,B = part− of

Background(π1, π2)

• SDRT provides an exact formulation of inferences in discourse interpretation

• nevertheless, world knowledge has to be encoded in a suitable fashion in order to matchthe antecedents of glue logic defaults

• it seems rather difficult to straightforwardly extend the approach to arbitrary non-domain-specific naturally-occuring texts

• ways to express (at least parts of) domain and world knowledge in a constrained manner:

- either use an extended view of lexical knowledge: Bos et al. (1995)

- or use frame-based representations of stereotypical knowledge

• proposal: use Frame Semantics developed by Fillmore (1976) and subsequent work onFrameNet (Baker et al., 1998) to get clues for the resolution of bridging anaphora

2Availability in SDRT says in essence that an antecedent for an anaphoric expression must be DRS-accessibleon the right frontier.

12

3 Frame Semantics and FrameNet

3.1 Basic Units

Frame Semantics: Basic Idea

• central assumption: world knowledge is organized in frames

• basic units: frames and lexical units

Frame : mental representation of a stereotypical situation whose elements can only bedefined relating one to another

Lexical Unit : a pairing of a word with a meaning; polysemous words are represented byseveral lexical units

• every lexical unit evokes a particular frame and can only be understood in relation tothat frame

FrameNet

• a lexical resource providing a body of annotated sentences based on frame semantics (Bakeret al., 1998; Ruppenhofer et al., 2005)

• the database contains around 10,000 lexical units, 800 semantic frames and over 120,000example sentences

Def. 1. Frame A frame is a structure 〈N,D,FE,LU〉 consisting of a name N , a definition D,a set of frame elements FE, and a set of lexical units LU .

Lexical Units

Def. 2. Lexical Unit Let Φ be a frame and V a linguistic expression that is potentially frame-evoking, i.e. a verb, a noun, or an adjective3, so

V ∈ LU(Φ) iff V evokes Φ.

3.2 Frame Elements

• FEs are kinds of entities that can participate in a frame

• core FEs: correspond roughly to thematic roles in an eventuality

• non-core FEs: include peripherical FEs and extra-thematic FEs

- peripheral FEs, e.g. Place, Time, Manner, Means, Degree can occur in any frame inwhich these FEs are semantically appropriate

- extra-thematic FEs introduce additional events that do not conceptually belong tothe frames they appear in

• example: the Killing frame4

Definition: A Killer or Cause causes the death of the Victim.

3Lexical units can also consist of several words, e.g. in the case of phrasal verbs such as “take a bath”.4Definitions are taken from the FrameNet Database, obtainable from the International Computer Science Insti-

tute, Berkeley, California (http://framenet.icsi.berkeley.edu/).

13

Core Frame Elements :

FE description inheritedFE

semantic type

Killer The person or sentient entity that causesthe death of the Victim

Agent sentient

Victim The living entity that dies as a result of thekilling

Patient sentient

Instrument The device used by the Killer to bringabout the death of the Victim

Instrument physical entity

Cause An inanimate entity or process that causesthe death of the Victim

Cause

Means The method or action that the Killer orCause performs resulting in the death ofthe Victim

Means state of affairs

Non-Core Frame Elements: Beneficiary, Manner, Place, Purpose, Time, ...Lexical Units: annihilate.v, annihilation.n, ..., murder.n, murder.v, murderer.n, ..., termi-

nate.v

• the lexical unit murder.v (which evokes the Killing frame)

Lexical Entry: murder.v– Frame elements and their syntactic realizations

Killer CNI.– (3), NP.Ext (15), PP[by].Dep (5)Victim NP.Ext (8), INI.– (1), NP.Obj(14)

– Frame elements and valence patterns

frame element realized asKiller NP.Ext NP.Ext PP[by].Dep CNI.–Victim NP.Obj INI.– NP.Ext NP.Ext(23) (14) (1) (5) (3)

Null Instantiations

• sometimes, conceptually necessary Frame Elements don’t show up in a sentence:

Constructional Null Instantiation, CNI : omitted agents in passive sentences(36) It informed him in letters an inch high that [V ictimCaptain Peter Dawson] had

been murderedTarget in Cyprus [KillerCNI]. (FrameNet, sentence 1761950)

(37) [V ictimJohn] was murderedTarget yesterday [KillerCNI].

Definite Null Instantiation, DNI (also anaphoric implicit FEs): missing obligatory ele-ments inferrable from the context

(38) [ThemeA tray of coffee] arrivedTarget almost immediately, together with an enor-mous plate of calorie-laden cakes [GoalDNI]. (FrameNet, sentence

999633)

Indefinite Null Instantiation, INI (also existential implicit FEs): implicit arguments ofcertain transitive verbs that are used intransitively (e.g. verbs as eat, bake)

(39) He had robbed, [Killerhe] had murderedTarget! [V ictimINI] (FrameNet, sentence

1762446)

14

3.3 Relations between Frames

Frame Inheritance

• frames are hierarchically organized: e.g. the frame Killing inherits the properties fromthe more general frame Transitive action which in turn inherits from the abstract frameEvent

• two types of frame inheritance are assumed: full and monotonic inheritance

Def. 3. Full Inheritance Let A and B be frames, so

inherits full(A,B) iff ∀ψ∃φ[ψ ∈ FE(B)→ φ ∈ FE(A)].

• for inherits full(A,B) to hold, the names of corresponding frame elements φ and ψ canbe different, but there must be some kind of binding between them

• the child frame A can have additional frame elements that are not part of B

• the relation inherits full has the following property: if A inherits from B then everyinstance of A is an instance of B

• multiple inheritance is explicitly allowed: a child frame can have multiple parents

• example: inheritance relations centered around the Killing frame 5

Def. 4. Monotonic Inheritance Let α and β be frames or frame elements, so

inherits monotonic(α, β) iff α v β.

• monotonic inheritance makes use of a sort6 hierarchy based on mereological relations likev (subtype)

• the sort hierarchy applies to frames, frame elements, and lexical units: α v β holds iff thesort of α is a subtype of sort β

5This representation is provided by FrameNet’s FrameGrapher application, which has been created by ChristineAdell Hodges and Carol Hays and is available online under http://framenet.icsi.berkeley.edu/FrameGrapher/.

6In FrameNet, ontological sorts are called “semantic types”. I will avoid this term and stick to the term “sort”in order to avoid confusion with what is usually understood as a semantic type, viz. entities e, truth-values t,relations from entities to truth-values 〈e, t〉 etc.

15

Uses

• the relation uses between frames is a kind of weak or partial inheritance

• if frame A uses frame B then frame A has FEs corresponding to FEs of B but need nothave an FE corresponding to each FE of B

• used FE must have compatible sorts

Def. 5. Uses If A, B are frames and uses(A,B) then

∃φ∃ψ[φ ∈ FE(A) ∧ ψ ∈ FE(B) ∧ inherits monotonic(φ, ψ)].

• for example, the frame Food uses the frame Ingestion implying an inheritance relationbetween the FEs Food in the former and Ingestibles in the latter frame

• Uses(A,B) is the set of used relations between FEs, viz. a two-place function from aframe A and a frame B to a subset of FE(A)× FE(B)

• for instance, Uses(Food, Ingestion) = {〈Food, Ingestibles〉}

Def. 6. Partial Inheritance If A and B are frames and uses(A,B), so

Uses(A,B) = U ⊆ FE(A)× FE(B) = {〈φ, ψ〉|φ ∈ FE(A) ∧ ψ ∈ FE(B)}.

Subframes

• many frames express complex events which may be constituted by several subevents

• for instance, the frame Crime scenario has the subframes Committing crime, Criminal investigation,and Criminal process

• subframes can contain FEs that are bound to FEs in the complex frame, e.g. the subframeCommitting crime has an FE Perpetrator, which is bound to the FE Perpetrator in thecomplex frame Crime scenario

• in contrast to inheritance, not all FEs of the superframe must necessarily show up in thesubframe

• often, subframe relations stand in a particular temporal ordering, which is specified by aseparate precedence relation, precedes

• if A is a subframe of B then an instance of A implies

1. the existence of an instance of B, and2. that all subframes of B which precede A are instantiated

• “e : Φ” stands for “e instantiates Φ”, where e denotes an eventuality

Def. 7. Subframe If A and B are frames and subframe(A,B) and e1 : A, then

∃e2[e2 : B], and

∀A′∃e′[subframe(A′, B) ∧ precedes(A′, A)→ e′ : A′].

16

3.4 Relations between Frame Elements

• relations between FEs in different frames:

- full inheritance of frames implies that corresponding FEs are inherited, too

- the partial inheritance relation uses also involves inheritance of FEs

• the relation inherits is the FE-to-FE relation corresponding to inheritance between frames,and inherits∗ is its transitive closure

Def. 8. Frame Element Inheritance Let A, B, and C be frames and φ ∈ FE(A) and ψ ∈FE(B), so

1. inherits(φ, ψ) iff

a. inherits full(A,B), orb. 〈φ, ψ〉 ∈ Uses(A,B).

2. inherits∗(φ, ψ) iff

a. inherits(φ, ψ) , orb. ∃χ[χ ∈ FE(C) ∧ inherits∗(φ, χ) ∧ inherits(χ, ψ)].

• in short, a FE φ in a frame A inherits from a FE ψ in another frame B if there is eitherfull inheritance between A and B, or if A uses B and φ is used as ψ

4 Marrying SDRT with FrameNet

4.1 Proposal: Integrate FrameNet and SDRT

• each eventuality introduced in a discourse evokes a corresponding frame in the discoursemodel

17

• its frame elements correspond to all relevant (necessary or optional) thematic roles of theevent

• proposal: include for all core frame elements a representation in the discourse model, i.e.in the SDRS of the current utterance

• in case that some participant of a frame is not expressed linguistically, its representationremains underspecified

• these elements can be further specified by subsequent information provided that the dis-course referent for the eventuality remains accessible for anaphoric reference

4.2 Representing Frame Elements in SDRT

• Neo-Davidsonian style of event semantics (Parsons, 1990)

- verbs include an implicit eventive argument in their semantic representation

- thematic roles in an eventuality are represented as conditions in form of predicateswith a first argument which is this eventive argument

- semantic representation of the sentence “John eats an apple.”: ∃e∃j∃a[eat(e) ∧agent(e, j) ∧ theme(e, a) ∧ named(j, john) ∧ apple(a)]

• the DRS construction rule for verbs has to be modified accordingly

• in the DRT implementation developed by Johan Bos (Curran et al., 2007), (6) is repre-sented by

(6) John was murdered yesterday.

(40)

e, j

named(j, john)murder(e)patient(e, j)

• How do frames enter the semantic representation?

• if a verb [V V], e.g. “murder”, introduces a condition V (e) in a DRS, together with adiscourse referent e for the corresponding eventuality, then the frame to which the verb ismapped by the corresponding lexical unit is evoked

• entities corresponding to frame elements show up as discourse referents xi

• the thematic role played by such an entity7 is represented as a relation φ between thediscourse referent xi and the discourse referent e corresponding to the eventuality, viz.φ(e, xi)

• two different kinds of discourse entities:

- regular discourse referents introduced by linguistic expressions, and

- weak discourse referents which are not (yet) expressed linguistically

7In fact, the term frame element as used in FrameNet is ambiguous: it can refer either to the entity that playsthe role corresponding to the frame element, i.e. the role-filler, or to the relation this entity has with respectto the frame.

18

• Weak Discourse Referents

- are abstract entities which are evoked or activated in course of interpretation

- a linguistic expression does not introduce them directly, but rather indirectly byvirtue of the frame evoked by a lexical unit

- often remain underspecified, but can be specified by subsequent anaphoric reference

• the universe of a DRS K is divided into two sets of referents:

– U rK : regular discourse referents, and– UwK : weak discourse referents,

with UK = U rK ∪ UwK

(41)

U rK | UwK

CK

Def. 9. Frame EvocationIf V ∈ LU(Φ) and e ∈ UK and V (e) ∈ CK then

∀ψ[ψ ∈ FE(Φ)→ ∃xi[xi ∈ UwK ∧ ψ(e, xi) ∈ CK ]] and

e : Φ ∈ CK .

(42) a.

e

murder(e)b.

e | x1, x2, x3

murder(e)e : Killingkiller(e, x1)victim(e, x2)instrument(e, x3)

• due to the inheritance between frames, the Killing frame inherits the frame elements fromthe parent frame Transitive Action

• the frame element Victim of the Killing frame is a specification of the frame elementPatient in the parent frame

• in order to avoid duplifications of DRS-conditions, a transition from a DRS K to a DRSK ′ matching inherited frame elements is defined

Def. 10. Matching Inherited Frame ElementsIf x ∈ UwK ∧ φ(e, x) ∈ CK then

a. y ∈ U rK′ ∧ φ(e, y) ∈ CK′ if y ∈ U rK ∧ ψ(e, y) ∈ CK and inherits∗(φ, ψ),b. x ∈ UwK′ ∧ φ(e, x) ∈ CK′ ∧ x =? ∈ CK′ otherwise.

(43) a. John was murdered yesterday.

b. The knife lay nearby.

19

(44)

π1, π2 | R, v

π1 :

e1, j | x1, x3

murder(e1)e1 : Killingkiller(e1, x1), x1 =?,victim(e1, j), named(j, john)instrument(e1, x3), x3 =?

π2 :

e2, !k | B, a

lie nearby(e2)e2 : Being locatedtheme(e2, k)knife(k)B(a, k), B =?, a =?

R(v, π2), R =?, v =?

4.3 Establishing Discourse Relations by FrameNet Data

4.3.1 Inferring Background

• the FrameNet database is hierarchically structured

- the Killing frame inherits the properties of Transitive action which in turn inheritsfrom Event

- the frame Being located inherits the frame elements of the abstract frame State

• Asher and Lascarides (2003): the occurrence of an event followed by a state is a strongindicator for the presence of a Background relation8 (cf. Asher and Lascarides, 2003, p.207, Vieu and Prevot, 2004, p. 486)

Background (?(α, β, λ) ∧ event(eα) ∧ state(eβ)) > Background(α, β, λ)

• in example (43), a Background relation R between π1 and π2 can be assumed

4.3.2 Inferring Narration

• Narration can defeasibly be inferred if an event provides an occasion for another:

Narration (?(α, β, λ) ∧ occasion(eα, eβ)) > Narration(α, β, λ)

• occasion can normally be inferred “if two event types of a certain kind are to be related.”(Asher and Lascarides, 2003, p. 201)

(45) a. Last April he claimed that more than $750,000 in cash was stolen from his Jerusalemresidence.

b. When police investigated the theft, Capucci insisted that the money had been returned,and asked them to drop the matter. (TIME Magazine, Sep. 02, 1974)

• steal.v evokes the frame Theft, which inherits from Committing crime

• investigate.v evokes the frame Criminal investigation

• both Committing crime and Criminal investigation are subframes of Crime scenario, whereCommitting crime precedes Criminal investigation

8’>’ is a nonmonotonic conditional operator. A > B means: if A then normally B.

20

• hence, there is reason to suspect that the event corresponding to Committing crime pro-vides an occasion for the eventuality corresponding to Criminal investigation

• more generally, if Φ1 and Φ2 are both subframes of a third frame Ψ and precede each otherthen normally an occasion relation holds between the two eventualities corresponding tothe frames

Def. 11. Inferring Occasion from Subframes

If Φ1, Φ2, Ψ are frames with e1 : Φ1 and e2 : Φ2 then

(subframe(Φ1,Ψ) ∧ subframe(Φ2,Ψ) ∧ precedes(Φ1,Φ2)) > occasion(e1, e2).

5 Constraints on Bridging Inferences

5.1 Problems for Resolving Bridging Anaphora

• in general, two problems are to be solved in bridging anaphora resolution:

1. find the correct antecedent to which the anaphor is to be connected– possible antecedents must be identified– impossible antecedents must be ruled out

2. identify the nature of the bridging relation itself– is it identity?– is it a mereological relation?– is it frame-related?

• Proposal for problem 1

– restrict the search space for suitable antecedents to take into account only accessibleregular and weak discourse referents

– accessibility for anaphoric reference is constrained by general discourse principles

• Proposal for problem 2

– restrict possible specifications of bridging relations to conditions on (weak or regular)discourse referents already present in the discourse model

5.2 Bridging Constraints

• Asher and Lascarides (1998a)’ meta-rules for bridging:

1. If possible use identity.2. Bridges must be plausible.3. Discourse structure determines bridging.4. Maximize discourse coherence.

5.2.1 The Preference for Coreference

• meta-rule 1 expresses the empirical preference that the referent of an anaphoric expressionis identical to another previously introduced referent

• seems to be subsumed by a very general constraint in discourse interpretation: DOAP(“Don’t overlook anaphoric possibilities”, (Williams, 1997))

• can be expressed by Equality by Default (Cohen, 2007): unless there is evidence for thecontrary, two (semantically compatible) discourse referents can be assumed to be equal

21

• accounts well for the resolution of direct anaphora

• in order to extend it to cover bridging anaphora, it must be adapted

• if a is an underspecified weak discourse referent in πj , then it wants to be resolved to anavailable discourse referent α introduced earlier

- for direct anaphora, α must be a regular discourse referent

- for indirect or bridging anaphora, α is a weak discourse referent

Def. 12. Equality by Default for Discourse Anaphora

a ∈ UwKπj ∧ α ∈ UKπi > a = α

• this default is to be understood as a last-resort principle; it can (and often must) be blockedby higher-ranked constraints in order to avoid overgeneration

5.2.2 The Plausibility Constraint

(46) a. John was murdered yesterday. b. # The book lay nearby.

• a book as killing instrument is less plausible, but still possible

• the lexical unit knife evokes the frame Weapon with the semantic type “artifact”: it ismore plausible

• knives are not necessarily designed as weapons

• the only usable knowledge: the frame element instrument must have a sort “physi-cal entity”: it is consistent with the sort of both lexical units book and knife: “artifact”

• plausibility entails consistence (Zeevat, 2006), but cannot be reduced to it: partly anapproximation of meta-rule 2

• some sort of probabilistic notion of plausibility is needed

5.2.3 The Right Frontier Constraint

• the Right Frontier of a tree are all nodes on the right extremeR0

R1

u1 u2

R2

u3 u4

• a coordinating discourse relation pushes the right frontier to the right, closing off its at-tachment point

• a subordinating discourse relation extends the right frontier downwards and leaves openits attachment point

• in SDRT: “an antecedent for an anaphoric expression must be DRS-accessible on the rightfrontier” (Asher & Lascarides 2003 Asher and Lascarides (2003))

• Vieu and Prevot (2004): Background should be considered as subordinating by default

• in (43), π1 lies on the right frontier of the discourse, and e1 is accessible for anaphoricreference in π2

• the discourse structure says that, in principle, a bridging relation can be established

• this covers Asher & Lascarides’ meta-rule 3

22

5.2.4 Maximize Discourse Coherence

• with the presence of a discourse relation between π1 and π2, the discourse referents in π1

are available for anaphoric reference in π2

• with Equality by Default, it can be assumed that a is equal to e1→ the bridging relation B(a, k) can be specified as instrument(e1, k)

• as a byproduct, the underspecified variable x3 in the condition instrument(e1, x3) in π1

can be resolved to k, yielding that instrument and knife refer to the same entity

• why should we do so?

• if the bridging relation can be resolved, the discourse turns out to be more coherent:intuition behind meta-rule 4 (MDC)

(47)

π1, π2

π1 :

e1, j | x1, x3

e1 : Killingkiller(e1, x1), x1 =?victim(e1, j), named(j, john)instrument(e1, x3)

π2 :

e2, !k | B, a

e2 : Being locatedtheme(e2, k), knife(k)B(a, k), B = instrument,a = e1, x3 = k

Background(π1, π2)

5.3 Summary

• to summarize the principles we need for bridging resolution, we remain with the followinggeneral constraints on anaphoric reference:

• - DOAP (via Equality by Default)

• - PLAUSIBLE (subsumes CONSISTENT)

• - RFC

• - MDC (minimize labels, maximize rhetorical connections, resolve underspecifications)

• they are not meant to be special meta-rules designed for bridging resolution, they ratherseem to be more general constraints to be obeyed in discourse interpretation

• more formally:

Def. 13. Constraints on Bridging Anaphora ResolutionIf x ∈ U rKπj and B, a ∈ UwKπj and B(a, x) ∈ CKπj and α ∈ UKπi thenby default, a = α holds for the following cases: (DOAP)

1. if α ∈ U rKπi then B = identity with x = a,2. if α ∈ UwKπi then either

a. B =v+ with x v+ α, orb. B = φ with φ(α, x) ∈ CKπi ,

provided that

• the sort of α is compatible to the sort of a, (CONSISTENT)• α is available for anaphoric conditions in πj, and (RFC)• MDC is not violated. (MDC)

23

Conclusion

We have ...

• ... made a distinction of two kinds of bridging anaphora

• ... spelled out how world knowledge (represented in frames) contributes to interpretation

• ... made a distinction between two types of discourse referents

• ... considerably constrained the resolution of bridging anaphora

In particular:

• only accessible regular and weak discourse referents are suitable bridging anchors• preferred bridging relations are conditions on discourse referents already present in the

discourse model• both indirect anaphora (involving roles in eventualities as well as mereological relations)

and direct anaphora are covered by this assumption

References

Asher, Nicholas and Alex Lascarides (1998a): ‘Bridging.’ Journal of Semantics, 15(1), 83–113.Asher, Nicholas and Alex Lascarides (1998b): ‘The semantics and pragmatics of presupposition.’ Journal of

Semantics, 15(2), 239–299.Asher, Nicholas and Alex Lascarides (2003): Logics of Conversation. Cambridge University Press, Cambridge.Baker, Collin, Charles Fillmore, and John Lowe (1998): ‘The Berkeley FrameNet project.’ In Christian Boitet

and Pete Whitelock, eds., Proceedings of the 36th ACL and 17th COLING, pp. 86–90. San Francisco, CA.Bos, Johan, Paul Buitelaar, and Anne-Marie Mineur (1995): ‘Bridging as coercive accommodation.’ In

Proceedings to the Workshop on Computational Logic for Natural Language Processing (CLNLP). Edinburgh.Chierchia, Gennaro (1995): Dynamics of Meaning: Anaphora, Presuppositions, and the Theory of Grammar.

University of Chicago Press, Chicago.Clark, Herbert (1977): ‘Bridging.’ In P. N. Johnson-Laird and P. Wason, eds., Thinking: Readings in Cognitive

Science, pp. 411–420. Cambridge.Cohen, Ariel (2007): ‘Anaphora resolution by default.’ In J. Geertzen, E. Thijsse, H. Bunt, and A. Schiffrin,

eds., IWCS-7, pp. 53–64. Tilburg, The Netherlands.Curran, James, Stephen Clark, , and Johan Bos (2007): ‘Linguistically motivated large-scale NLP with C&C

and Boxer.’ In Proceedings of the ACL 2007 Demonstrations Session (ACL-07 demo), pp. 29–32.Fillmore, Charles (1976): ‘Frame semantics and the nature of language.’ In Annals of the NY Academy of

Sciences: Conf. on the Origin and Development of Language and Speech 280, pp. 20–32.Fraurud, Kari (1990): ‘Definiteness and the processing of NPs in natural discourse.’ Journal of Semantics, 7,

395–433.Gardent, Claire, Helene Manuelian, and Eric Kow (2003): ‘Which bridges for bridging definite descriptions?’

In Proceedings of LINC’03. Budapest, Hungary.Garrod, Simon and Melody Terras (2000): ‘The contribution of lexical and situational knowledge to resolving

discourse roles: Bonding and resolution.’ Journal of Memory and Language, 42, 526–544.Kamp, Hans and Uwe Reyle (1993): From Discourse to Logic. Introduction to Modeltheoretic Semantics of

Natural Language, Formal Logic and Discourse Representation Theory. Kluwer, Dordrecht.Parsons, Terence (1990): Events in the Semantics of English. MIT Press, Cambridge, MA.Poesio, Massimo and Renata Vieira (1998): ‘A corpus-based investigation of definite descriptions use.’ Compu-

tational Linguistics, 24(2), 183–216.Prince, Ellen (1981): ‘Towards a taxonomy of given-new information.’ In Peter Cole, ed., Radical Pragmatics,

pp. 223–256. Academic Press, New York.Ruppenhofer, Josef, Michael Ellsworth, Miriam Petruck, Christopher Johnson, and Jan Scheffczyk

(2005): ‘FrameNet II: Extended theory and practice.’ Technical report, ICSI, Berkeley.Russell, Bertrand (1905): ‘On Denoting.’ Mind, 14, 479–493.Sanford, A. J., Simon Garrod, A. Lucas, and R. Henderson (1993): ‘Pronouns without explicit antecedents?’

Journal of Semantics, 2(3/4), 303–318.Schwarz, Monika (2000): Indirekte Anaphern in Texten: Studien zur domanengebundenen Referenz und Koharenz

im Deutschen. Niemeyer, Tuebingen.Schwarz-Friesel, Monika (2007): ‘Indirect anaphora in text. A cognitive account.’ In Monika Schwarz-Friesel,

Manfred Consten, and Mareile Hillevi Knees, eds., Anaphors in Text, Studies in Language Companion Series86. John Benjamins, Philadelphia.

Vendler, Zeno (1957): ‘Verbs and times.’ Philosophical Review, 66(2), 143–160.Vieu, Laure and Laurent Prevot (2004): ‘Background in SDRT.’ In Workshop SDRT, Traitement Automatique

du Langage Naturel, Fez, Morocco, pp. 485–494.Williams, Edwin (1997): ‘Blocking and anaphora.’ Linguistic Inquiry, 28, 577–628.Zeevat, Henk (2006): ‘Discourse structure in optimality theoretic pragmatics.’ In John Harpur, Anton Benz,

Candace Sidner, and Peter Kuhnlein, eds., Proceedings of Constraints in Discourse II (CID-06), Maynooth,Ireland, pp. 155–161.

24