word sense disambiguation - semantic scholar 18th european summer school in logic, language and...

The 18th European Summer School in Logic, Language and Information

31 July - 11 August, 2006 � Málaga, Spain

Word Sense Disambiguation

Rada Mihalcea

Introductory Course, Language and Computation Section

ESSLLI is the Annual Summer School of FoLLI,

The Association for Logic, Language and Information

http://www.folli.org

Organized by Sponsored by

1

Word Sense DisambiguationTutorial at ESSLLI 2006

Rada MihalceaUniversity of North Texas

http://www.cs.unt.edu/~rada

Note: these tutorial notes were created in collaboration with Ted Pedersen, University of Minnesota, Duluth, http://www.d.umn.edu/~tpederse/

2

Goal of the Tutorial

• Introduce the problem of word sense disambiguation (WSD), focusing on the range of formulations and approaches currently practiced.

• Accessible to anyone with an interest in Artificial Intelligence and NLP.

• Persuade you to work on word sense disambiguation– It is an interesting problem– Lots of good work already done, still more to do– There is infrastructure to help you get started

• Persuade you to use word sense disambiguation in your text applications.

2

3

Outline of Tutorial

• Introduction • Methodolodgy• Knowledge Intensive Methods• Supervised Approaches• Minimally Supervised Approaches• Unsupervised Learning• How to Get Started• Conclusion

Part 1:Introduction

3

5

Outline

• Definitions• Ambiguity for Humans and Computers• Very Brief Historical Overview• Theoretical Connections• Practical Applications

6

Definitions

• Word sense disambiguation is the problem of selecting a sense for a word from a set of predefined possibilities. – Sense Inventory usually comes from a dictionary or thesaurus.– Knowledge intensive methods, supervised learning, and

(sometimes) bootstrapping approaches

• Word sense discrimination is the problem of dividing the usages of a word into different meanings, without regard to any particular existing sense inventory.– Unsupervised techniques

4

7

Computers versus Humans

• Polysemy – most words have many possible meanings.• A computer program has no basis for knowing which one

is appropriate, even if it is obvious to a human…• Ambiguity is rarely a problem for humans in their day to

day communication, except in extreme cases…

8

Ambiguity for Humans - Newspaper Headlines!

• DRUNK GETS NINE YEARS IN VIOLIN CASE• FARMER BILL DIES IN HOUSE • PROSTITUTES APPEAL TO POPE • STOLEN PAINTING FOUND BY TREE • RED TAPE HOLDS UP NEW BRIDGE• DEER KILL 300,000• RESIDENTS CAN DROP OFF TREES• INCLUDE CHILDREN WHEN BAKING COOKIES • MINERS REFUSE TO WORK AFTER DEATH

5

9

Ambiguity for a Computer

• The fisherman jumped off the bank and into the water.• The bank down the street was robbed!• Back in the day, we had an entire bank of computers

devoted to this problem. • The bank in that road is entirely too steep and is really

dangerous. • The plane took a bank to the left, and then headed off

towards the mountains.

10

Early Days of WSD

• Noted as problem for Machine Translation (Weaver, 1949)– A word can often only be translated if you know the specific sense

intended (A bill in English could be a pico or a cuenta in Spanish)

• Bar-Hillel (1960) posed the following:– Little John was looking for his toy box. Finally, he found it. The

box was in the pen. John was very happy.– Is “pen” a writing instrument or an enclosure where children play?

…declared it unsolvable, left the field of MT!

6

11

Since then…

• 1970s - 1980s – Rule based systems– Rely on hand crafted knowledge sources

• 1990s – Corpus based approaches– Dependence on sense tagged text– (Ide and Veronis, 1998) overview history from early days to 1998.

• 2000s – Hybrid Systems– Minimizing or eliminating use of sense tagged text– Taking advantage of the Web

12

Theoretical Connections

• Cognitive Science & Psychology– Quillian (1968), Collins and Loftus (1975) : spreading activation

• Hirst (1987) developed marker passing model

• Linguistics – Fodor & Katz (1963) : selectional preferences

• Resnik (1993) pursued statistically

• Philosophy of Language– Wittgenstein (1958): meaning as use – “For a large class of cases-though not for all-in which we employ

the word "meaning" it can be defined thus: the meaning of a wordis its use in the language.”

7

13

Practical Applications

• Machine Translation– Translate “bill” from English to Spanish

• Is it a “pico” or a “cuenta”?• Is it a bird jaw or an invoice?

• Information Retrieval– Find all Web Pages about “cricket”

• The sport or the insect?

• Question Answering– What is George Miller’s position on gun control?

• The psychologist or US congressman?

• Knowledge Acquisition– Add to KB: Herb Bergson is the mayor of Duluth.

• Minnesota or Georgia?

14

References

• (Bar-Hillel, 1960) The Present Status of Automatic Translations of Languages. In Advances in Computers. Volume 1. Alt, F. (editor). Academic Press, New York, NY. pp 91-163.

• (Collins and Loftus, 1975) A Spreading Activation Theory of Semantic Memory. Psychological Review, (82) pp. 407-428.

• (Fodor and Katz, 1963) The structure of semantic theory. Language (39). pp 170-210. • (Hirst, 1987) Semantic Interpretation and the Resolution of Ambiguity. Cambridge

University Press. • (Ide and Véronis, 1998)Word Sense Disambiguation: The State of the Art..

Computational Linguistics (24) pp 1-40.• (Quillian, 1968) Semantic Memory. In Semantic Information Processing. Minsky, M.

(editor). The MIT Press, Cambridge, MA. pp. 227-270. • (Resnik, 1993) Selection and Information: A Class-Based Approach to Lexical

Relationships. Ph.D. Dissertation. University of Pennsylvania. • (Weaver, 1949): Translation. In Machine Translation of Languages: fourteen essays.

Locke, W.N. and Booth, A.D. (editors) The MIT Press, Cambridge, Mass. pp. 15-23. • (Wittgenstein, 1958) Philosophical Investigations, 3rd edition. Translated by G.E.M.

Anscombe. Macmillan Publishing Co., New York.

8

Part 2:Methodology

16

Outline

• General considerations• All-words disambiguation• Targeted-words disambiguation• Word sense discrimination, sense discovery• Evaluation (granularity, scoring)

9

17

–Ex: “chair” – furniture or person–Ex: “child” – young person or human offspring

Overview of the Problem• Many words have several meanings (homonymy / polysemy)

• Determine which sense of a word is used in a specific sentence

• Note:– often, the different senses of a word are closely related

• Ex: title - right of legal ownership- document that is evidence of the legal ownership,

– sometimes, several senses can be “activated” in a single context (co-activation)

• Ex: “This could bring competition to the trade”competition: - the act of competing

- the people who are competing

18

Word Senses

• The meaning of a word in a given context

• Word sense representations– With respect to a dictionary

chair = a seat for one person, with a support for the back; "he put his coat over the back of the chair and sat down"chair = the position of professor; "he was awarded an endowed chair ineconomics"

– With respect to the translation in a second languagechair = chaisechair = directeur

– With respect to the context where it occurs (discrimination)“Sit on a chair” “Take a seat on this chair”“The chair of the Math Department” “The chair of the meeting”

10

19

Approaches to Word Sense Disambiguation

• Knowledge-Based Disambiguation– use of external lexical resources such as dictionaries and thesauri– discourse properties

• Supervised Disambiguation– based on a labeled training set– the learning system has:

• a training set of feature-encoded inputs AND • their appropriate sense label (category)

• Unsupervised Disambiguation– based on unlabeled corpora– The learning system has:

• a training set of feature-encoded inputs BUT • NOT their appropriate sense label (category)

20

All Words Word Sense Disambiguation

• Attempt to disambiguate all open-class words in a text“He put his suit over the back of the chair”

• Knowledge-based approaches• Use information from dictionaries

– Definitions / Examples for each meaning• Find similarity between definitions and current context

• Position in a semantic network• Find that “table” is closer to “chair/furniture” than to “chair/person”

• Use discourse properties• A word exhibits the same sense in a discourse / in a collocation

11

21

All Words Word Sense Disambiguation

• Minimally supervised approaches– Learn to disambiguate words using small annotated corpora– E.g. SemCor – corpus where all open class words are

disambiguated• 200,000 running words

• Most frequent sense

22

Targeted Word Sense Disambiguation

• Disambiguate one target word“Take a seat on this chair”“The chair of the Math Department”

• WSD is viewed as a typical classification problem– use machine learning techniques to train a system

• Training:– Corpus of occurrences of the target word, each occurrence

annotated with appropriate sense– Build feature vectors:

• a vector of relevant linguistic features that represents the context (ex: a window of words around the target word)

• Disambiguation:– Disambiguate the target word in new unseen text

12

23

Targeted Word Sense Disambiguation

• Take a window of n word around the target word• Encode information about the words around the target word

– typical features include: words, root forms, POS tags, frequency, …• An electric guitar and bass player stand off to one side, not really part of

the scene, just as a sort of nod to gringo expectations perhaps.

• Surrounding context (local features)– [ (guitar, NN1), (and, CJC), (player, NN1), (stand, VVB) ]

• Frequent co-occurring words (topical features)– [fishing, big, sound, player, fly, rod, pound, double, runs, playing, guitar, band]– [0,0,0,1,0,0,0,0,0,0,1,0]

• Other features:– [followed by "player", contains "show" in the sentence,…] – [yes, no, … ]

24

Unsupervised Disambiguation

• Disambiguate word senses:– without supporting tools such as dictionaries and thesauri – without a labeled training text

• Without such resources, word senses are not labeled– We cannot say “chair/furniture” or “chair/person”

• We can:– Cluster/group the contexts of an ambiguous word into a number

of groups – Discriminate between these groups without actually labeling

them

13

25

Unsupervised Disambiguation

• Hypothesis: same senses of words will have similar neighboring words

• Disambiguation algorithm– Identify context vectors corresponding to all occurrences of a particular

word – Partition them into regions of high density– Assign a sense to each such region

“Sit on a chair”“Take a seat on this chair”“The chair of the Math Department”“The chair of the meeting”

26

Evaluating Word Sense Disambiguation

• Metrics: – Precision = percentage of words that are tagged correctly, out of the

words addressed by the system– Recall = percentage of words that are tagged correctly, out of all words

in the test set– Example

• Test set of 100 words Precision = 50 / 75 = 0.66• System attempts 75 words Recall = 50 / 100 = 0.50• Words correctly disambiguated 50

• Special tags are possible:– Unknown– Proper noun– Multiple senses

• Compare to a gold standard – SEMCOR corpus, SENSEVAL corpus, …

14

27

Evaluating Word Sense Disambiguation

• Difficulty in evaluation:– Nature of the senses to distinguish has a huge impact on results

• Coarse versus fine-grained sense distinctionchair = a seat for one person, with a support for the back; "he put his coat

over the back of the chair and sat down“chair = the position of professor; "he was awarded an endowed chair in

economics“

bank = a financial institution that accepts deposits and channels the money into lending activities; "he cashed a check at the bank"; "that bank holds the mortgage on my home"

bank = a building in which commercial banking is transacted; "the bank is on the corner of Nassau and Witherspoon“

• Sense maps– Cluster similar senses– Allow for both fine-grained and coarse-grained evaluation

28

Bounds on Performance

• Upper and Lower Bounds on Performance: – Measure of how well an algorithm performs relative to the difficulty of

the task.

• Upper Bound: – Human performance– Around 97%-99% with few and clearly distinct senses– Inter-judge agreement:

• With words with clear & distinct senses – 95% and up• With polysemous words with related senses – 65% – 70%

• Lower Bound (or baseline): – The assignment of a random sense / the most frequent sense

• 90% is excellent for a word with 2 equiprobable senses• 90% is trivial for a word with 2 senses with probability ratios of 9 to 1

15

29

References• (Gale, Church and Yarowsky 1992) Gale, W., Church, K., and Yarowsky, D. Estimating

upper and lower bounds on the performance of word-sense disambiguation programs ACL 1992.

• (Miller et. al., 1994) Miller, G., Chodorow, M., Landes, S., Leacock, C., and Thomas, R. Using a semantic concordance for sense identification. ARPA Workshop 1994.

• (Miller, 1995) Miller, G. Wordnet: A lexical database. ACM, 38(11) 1995.• (Senseval) Senseval evaluation exercises http://www.senseval.org

Part 3:Knowledge-based Methods for Word Sense Disambiguation

16

31

Outline

• Task definition– Machine Readable Dictionaries

• Algorithms based on Machine Readable Dictionaries• Selectional Restrictions• Measures of Semantic Similarity• Heuristic-based Methods

32

Task Definition

• Knowledge-based WSD = class of WSD methods relying (mainly) on knowledge drawn from dictionaries and/or raw text

• Resources– Yes

• Machine Readable Dictionaries• Raw corpora

– No• Manually annotated corpora

• Scope– All open-class words

17

33

Machine Readable Dictionaries

• In recent years, most dictionaries made available in Machine Readable format (MRD)– Oxford English Dictionary– Collins– Longman Dictionary of Ordinary Contemporary English (LDOCE)

• Thesauruses – add synonymy information– Roget Thesaurus

• Semantic networks – add more semantic relations– WordNet– EuroWordNet

34

MRD – A Resource for Knowledge-based WSD• For each word in the language vocabulary, an MRD

provides:– A list of meanings– Definitions (for all word meanings)– Typical usage examples (for most word meanings)

WordNet definitions/examples for the noun plant1. buildings for carrying on industrial labor; "they built a large plant to

manufacture automobiles“2. a living organism lacking the power of locomotion3. something planted secretly for discovery by another; "the police used a plant to

trick the thieves"; "he claimed that the evidence against him was a plant"4. an actor situated in the audience whose acting is rehearsed but seems

spontaneous to the audience

18

35

MRD – A Resource for Knowledge-based WSD

• A thesaurus adds:– An explicit synonymy relation between word meanings

• A semantic network adds:– Hypernymy/hyponymy (IS-A), meronymy/holonymy (PART-OF),

antonymy, entailnment, etc.

WordNet synsets for the noun “plant”1. plant, works, industrial plant2. plant, flora, plant life

WordNet related concepts for the meaning “plant life”{plant, flora, plant life}

hypernym: {organism, being}hypomym: {house plant}, {fungus}, …meronym: {plant tissue}, {plant part}holonym: {Plantae, kingdom Plantae, plant kingdom}

36

Outline



19

37

Lesk Algorithm

• (Michael Lesk 1986): Identify senses of words in context using definition overlap

Algorithm:1. Retrieve from MRD all sense definitions of the words to be

disambiguated2. Determine the definition overlap for all possible sense combinations3. Choose senses that lead to highest overlap

Example: disambiguate PINE CONE• PINE

1. kinds of evergreen tree with needle-shaped leaves2. waste away through sorrow or illness

• CONE 1. solid body which narrows to a point2. something of this shape whether solid or hollow3. fruit of certain evergreen trees

Pine#1 ∩ Cone#1 = 0Pine#2 ∩ Cone#1 = 0Pine#1 ∩ Cone#2 = 1Pine#2 ∩ Cone#2 = 0Pine#1 ∩ Cone#3 = 2Pine#2 ∩ Cone#3 = 0

38

Lesk Algorithm for More than Two Words?

• I saw a man who is 98 years old and can still walk and tell jokes– nine open class words: see(26), man(11), year(4), old(8), can(5), still(4),

walk(10), tell(8), joke(3)

• 43,929,600 sense combinations! How to find the optimal sense combination?

• Simulated annealing (Cowie, Guthrie, Guthrie 1992)– Define a function E = combination of word senses in a given text.– Find the combination of senses that leads to highest definition overlap

(redundancy)1. Start with E = the most frequent sense for each word2. At each iteration, replace the sense of a random word in the set with a different sense, and measure E

3. Stop iterating when there is no change in the configuration of senses

20

39

Lesk Algorithm: A Simplified Version

• Original Lesk definition: measure overlap between sense definitions for all words in context– Identify simultaneously the correct senses for all words in context

• Simplified Lesk (Kilgarriff & Rosensweig 2000): measure overlap between sense definitions of a word and current context– Identify the correct sense for one word at a time

• Search space significantly reduced

40

Lesk Algorithm: A Simplified Version

Example: disambiguate PINE in

“Pine cones hanging in a tree”

• PINE

1. kinds of evergreen tree with needle-shaped leaves

2. waste away through sorrow or illness

Pine#1 ∩ Sentence = 1Pine#2 ∩ Sentence = 0

• Algorithm for simplified Lesk:1.Retrieve from MRD all sense definitions of the word to be disambiguated

2.Determine the overlap between each sense definition and the current context

3.Choose the sense that leads to highest overlap

21

41

Evaluations of Lesk Algorithm

• Initial evaluation by M. Lesk– 50-70% on short samples of text manually annotated set, with respect

to Oxford Advanced Learner’s Dictionary

• Simulated annealing – 47% on 50 manually annotated sentences

• Evaluation on Senseval-2 all-words data, with back-off to random sense (Mihalcea & Tarau 2004)– Original Lesk: 35%– Simplified Lesk: 47%

• Evaluation on Senseval-2 all-words data, with back-off to most frequent sense (Vasilescu, Langlais, Lapalme 2004)– Original Lesk: 42%– Simplified Lesk: 58%

42

Outline


• Algorithms based on Machine Readable Dictionaries• Selectional Preferences• Measures of Semantic Similarity• Heuristic-based Methods

22

43

Selectional Preferences

• A way to constrain the possible meanings of words in a given context

• E.g. “Wash a dish” vs. “Cook a dish”– WASH-OBJECT vs. COOK-FOOD

• Capture information about possible relations between semantic classes – Common sense knowledge

• Alternative terminology– Selectional Restrictions – Selectional Preferences– Selectional Constraints

44

Acquiring Selectional Preferences

• From annotated corpora– Circular relationship with the WSD problem

• Need WSD to build the annotated corpus• Need selectional preferences to derive WSD

• From raw corpora – Frequency counts– Information theory measures– Class-to-class relations

23

45

Preliminaries: Learning Word-to-Word Relations

• An indication of the semantic fit between two words1. Frequency counts

– Pairs of words connected by a syntactic relations

2. Conditional probabilities– Condition on one of the words

),,( 21 RWWCount

),( ),,( ),|(

2

2121 RWCount

RWWCountRWWP =

46

Learning Selectional Preferences (1)

• Word-to-class relations (Resnik 1993)– Quantify the contribution of a semantic class using all the concepts

subsumed by that class

– where

)(),|(log),|(

)(),|(log),|(

),,(

2

1212

2

1212

21

2CP

RWCPRWCP

CPRWCPRWCP

RCWA

C∑

=

∑∈

=

=

22)(

),,(),,(

),(

),,(),|(

2

2121

1

2112

CW WCountRWWCountRCWCount

RWCountRCWCountRWCP

24

47

Learning Selectional Preferences (2)• Determine the contribution of a word sense based on the assumption of

equal sense distributions:– e.g. “plant” has two senses 50% occurences are sense 1, 50% are sense 2

• Example: learning restrictions for the verb “to drink”– Find high-scoring verb-object pairs

– Find “prototypical” object classes (high association score)

Co-occ score Verb Object11.75 drink tea11.75 drink Pepsi11.75 drink champagne10.53 drink liquid10.2 drink beer9.34 drink wine

A(v,c) Object class3.58 (beverage, [drink, …])2.05 (alcoholic_beverage, [intoxicant, …])

48

Learning Selectional Preferences (3)

• Other algorithms– Learn class-to-class relations (Agirre and Martinez, 2002)

• E.g.: “ingest food” is a class-to-class relation for “eat chicken”

– Bayesian networks (Ciaramita and Johnson, 2000)– Tree cut model (Li and Abe, 1998)

25

49

Using Selectional Preferences for WSD

Algorithm:1. Learn a large set of selectional preferences for a given syntactic

relation R2. Given a pair of words W1– W2 connected by a relation R3. Find all selectional preferences W1– C (word-to-class) or C1– C2

(class-to-class) that apply4. Select the meanings of W1 and W2 based on the selected semantic

class

Example: disambiguate coffee in “drink coffee”1. (beverage) a beverage consisting of an infusion of ground coffee beans

2. (tree) any of several small trees native to the tropical Old World

3. (color) a medium to dark brown colorGiven the selectional preference “DRINK BEVERAGE” : coffee#1

50

Evaluation of Selectional Preferences for WSD

• Data set– mainly on verb-object, subject-verb relations extracted from

SemCor

• Compare against random baseline• Results (Agirre and Martinez, 2000)

– Average results on 8 nouns– Similar figures reported in (Resnik 1997)

Object SubjectPrecision Recall Precision Recall

Random 19.2 19.2 19.2 19.2Word-to-word 95.9 24.9 74.2 18.0Word-to-class 66.9 58.0 56.2 46.8Class-to-class 66.6 64.8 54.0 53.7

26

51

Outline



52

Semantic Similarity

• Words in a discourse must be related in meaning, for the discourse to be coherent (Haliday and Hassan, 1976)

• Use this property for WSD – Identify related meanings for words that share a common context

• Context span:1. Local context: semantic similarity between pairs of words2. Global context: lexical chains

27

53

Semantic Similarity in a Local Context

• Similarity determined between pairs of concepts, or between a word and its surrounding context

• Relies on similarity metrics on semantic networks– (Rada et al. 1989)

carnivore

wild dogwolf

bearfeline, felidcanine, canidfissiped mamal, fissiped

dachshund

hunting doghyena dogdingo

hyenadog

terrier

54

Semantic Similarity Metrics (1)• Input: two concepts (same part of speech) • Output: similarity measure• (Leacock and Chodorow 1998)

– E.g. Similarity(wolf,dog) = 0.60 Similarity(wolf,bear) = 0.42

• (Resnik 1995)– Define information content, P(C) = probability of seeing a concept of type

C in a large corpus

– Probability of seeing a concept = probability of seeing instances of that concept

– Determine the contribution of a word sense based on the assumption of equal sense distributions:

• e.g. “plant” has two senses 50% occurrences are sense 1, 50% are sense 2

−=

DCCPathCCSimilarity

2),(log),( 21

21 , D is the taxonomy depth

))(log()( CPCIC −=

28

55

Semantic Similarity Metrics (2)

• Similarity using information content– (Resnik 1995) Define similarity between two concepts (LCS = Least

Common Subsumer)

– Alternatives (Jiang and Conrath 1997)

• Other metrics:– Similarity using information content (Lin 1998)– Similarity using gloss-based paths across different hierarchies (Mihalcea

and Moldovan 1999)– Conceptual density measure between noun semantic hierarchies and

current context (Agirre and Rigau 1995)– Adapted Lesk algorithm (Banerjee and Pedersen 2002)

)),((),( 2121 CCLCSICCCSimilarity =

))()(( )),((2),(

21

2121

CICCICCCLCSICCCSimilarity

+−×=

56

Semantic Similarity Metrics for WSD

• Disambiguate target words based on similarity with one word to the left and one word to the right– (Patwardhan, Banerjee, Pedersen 2002)

• Evaluation:– 1,723 ambiguous nouns from Senseval-2– Among 5 similarity metrics, (Jiang and Conrath 1997) provide the

best precision (39%)

Example: disambiguate PLANT in “plant with flowers”PLANT1. plant, works, industrial plant2. plant, flora, plant life

Similarity (plant#1, flower) = 0.2Similarity (plant#2, flower) = 1.5 : plant#2

29

57

Semantic Similarity in a Global Context

• Lexical chains (Hirst and St-Onge 1988), (Haliday and Hassan 1976)

• “A lexical chain is a sequence of semantically related words, which creates a context and contributes to the continuity of meaning and the coherence of a discourse”

Algorithm for finding lexical chains:1. Select the candidate words from the text. These are words for which we can

compute similarity measures, and therefore most of the time they have the same part of speech.

2. For each such candidate word, and for each meaning for this word, find a chain to receive the candidate word sense, based on a semantic relatedness measure between the concepts that are already in the chain, and the candidate word meaning.

3. If such a chain is found, insert the word in this chain; otherwise, create a new chain.

58

Semantic Similarity of a Global ContextA very long train traveling along the rails with a constant velocity v in a certain direction …

train #1: public transport

#2: order set of things

#3: piece of cloth

travel

#1 change location

#2: undergo transportation

rail #1: a barrier

# 2: a bar of steel for trains

#3: a small bird

30

59

Lexical Chains for WSD

• Identify lexical chains in a text– Usually target one part of speech at a time

• Identify the meaning of words based on their membership to a lexical chain

• Evaluation:– (Galley and McKeown 2003) lexical chains on 74 SemCor texts

give 62.09%– (Mihalcea and Moldovan 2000) on five SemCor texts give 90%

with 60% recall• lexical chains “anchored” on monosemous words

– (Okumura and Honda 1994) lexical chains on five Japanese texts give 63.4%

60

Outline



31

61

Example: “plant/flora” is used more often than “plant/factory”- annotate any instance of PLANT as “plant/flora”

Most Frequent Sense (1)

• Identify the most often used meaning and use this meaning by default

• Word meanings exhibit a Zipfian distribution– E.g. distribution of word senses in SemCor

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 2 3 4 5 6 7 8 9 10

Sense number

Freq

uenc

y Noun

Verb

Adj

Adv

62

Most Frequent Sense (2)• Method 1: Find the most frequent sense in an annotated

corpus• Method 2: Find the most frequent sense using a method

based on distributional similarity (McCarthy et al. 2004)

1. Given a word w, find the top k distributionally similar words Nw = {n1, n2, …, nk}, with associated similarity scores {dss(w,n1), dss(w,n2), … dss(w,nk)}

2. For each sense wsi of w, identify the similarity with the words nj, using the sense of nj that maximizes this score

3. Rank senses wsi of w based on the total similarity score

)),((max),( where

,),'(

),(),()(

)(

)('

xinsensesnsji

Nnwsensesws

ji

jiji

nswswnssnwswnss

nwswnssnwswnss

nwdsswsScore

jx

wj

i

∈

∈∈

=

= ∑ ∑

32

63

Most Frequent Sense(3)

• Word senses– pipe #1 = tobacco pipe– pipe #2 = tube of metal or plastic

• Distributional similar words – N = {tube, cable, wire, tank, hole, cylinder, fitting, tap, …}

• For each word in N, find similarity with pipe#i (using the sense that maximizes the similarity)

– pipe#1 – tube (#3) = 0.3– pipe#2 – tube (#1) = 0.6

• Compute score for each sense pipe#i– score (pipe#1) = 0.25– score (pipe#2) = 0.73

Note: results depend on the corpus used to find distributionallysimilar words => can find domain specific predominant senses

64

E.g. The ambiguous word PLANT occurs 10 times in a discourse all instances of “plant” carry the same meaning

One Sense Per Discourse

• A word tends to preserve its meaning across all its occurrences in a given discourse (Gale, Church, Yarowksy 1992)

• What does this mean?

• Evaluation: – 8 words with two-way ambiguity, e.g. plant, crane, etc.– 98% of the two-word occurrences in the same discourse carry the same

meaning• The grain of salt: Performance depends on granularity

– (Krovetz 1998) experiments with words with more than two senses– Performance of “one sense per discourse” measured on SemCor is approx.

70%

33

65

The ambiguous word PLANT preserves its meaning in all its occurrences within the collocation “industrial plant”, regardless of the context where this collocation occurs

One Sense per Collocation

• A word tends to preserver its meaning when used in the same collocation (Yarowsky 1993)

– Strong for adjacent collocations– Weaker as the distance between words increases

• An example

• Evaluation:– 97% precision on words with two-way ambiguity

• Finer granularity:– (Martinez and Agirre 2000) tested the “one sense per collocation”

hypothesis on text annotated with WordNet senses – 70% precision on SemCor words

66

References• (Agirre and Rigau, 1995) Agirre, E. and Rigau, G. A proposal for word sense disambiguation

using conceptual distance. RANLP 1995.• (Agirre and Martinez 2001) Agirre, E. and Martinez, D. Learning class-to-class selectional

preferences. CONLL 2001.• (Banerjee and Pedersen 2002) Banerjee, S. and Pedersen, T. An adapted Lesk algorithm for

word sense disambiguation using WordNet. CICLING 2002.• (Cowie, Guthrie and Guthrie 1992), Cowie, L. and Guthrie, J. A. and Guthrie, L.: Lexical

disambiguation using simulated annealing. COLING 2002.• (Gale, Church and Yarowsky 1992) Gale, W., Church, K., and Yarowsky, D. One sense per

discourse. DARPA workshop 1992.• (Halliday and Hasan 1976) Halliday, M. and Hasan, R., (1976). Cohesion in English.

Longman.• (Galley and McKeown 2003) Galley, M. and McKeown, K. (2003) Improving word sense

disambiguation in lexical chaining. IJCAI 2003• (Hirst and St-Onge 1998) Hirst, G. and St-Onge, D. Lexical chains as representations of

context in the detection and correction of malaproprisms. WordNet: An electronic lexical database, MIT Press.

• (Jiang and Conrath 1997) Jiang, J. and Conrath, D. Semantic similarity based on corpus statistics and lexical taxonomy. COLING 1997.

• (Krovetz, 1998) Krovetz, R. More than one sense per discourse. ACL-SIGLEX 1998.• (Lesk, 1986) Lesk, M. Automatic sense disambiguation using machine readable dictionaries:

How to tell a pine cone from an ice cream cone. SIGDOC 1986.• (Lin 1998) Lin, D An information theoretic definition of similarity. ICML 1998.

34

67

References• (Martinez and Agirre 2000) Martinez, D. and Agirre, E. One sense per collocation and

genre/topic variations. EMNLP 2000.• (Miller et. al., 1994) Miller, G., Chodorow, M., Landes, S., Leacock, C., and Thomas, R.

Using a semantic concordance for sense identification. ARPA Workshop 1994.• (Miller, 1995) Miller, G. Wordnet: A lexical database. ACM, 38(11) 1995.• (Mihalcea and Moldovan, 1999) Mihalcea, R. and Moldovan, D. A method for word

sense disambiguation of unrestricted text. ACL 1999. • (Mihalcea and Moldovan 2000) Mihalcea, R. and Moldovan, D. An iterative approach

to word sense disambiguation. FLAIRS 2000.• (Mihalcea, Tarau, Figa 2004) R. Mihalcea, P. Tarau, E. Figa PageRank on Semantic

Networks with Application to Word Sense Disambiguation, COLING 2004.• (Patwardhan, Banerjee, and Pedersen 2003) Patwardhan, S. and Banerjee, S. and

Pedersen, T. Using Measures of Semantic Relatedeness for Word Sense Disambiguation. CICLING 2003.

• (Rada et al 1989) Rada, R. and Mili, H. and Bicknell, E. and Blettner, M. Development and application of a metric on semantic nets. IEEE Transactions on Systems, Man, and Cybernetics, 19(1) 1989.

• (Resnik 1993) Resnik, P. Selection and Information: A Class-Based Approach to Lexical Relationships. University of Pennsylvania 1993.

• (Resnik 1995) Resnik, P. Using information content to evaluate semantic similarity. IJCAI 1995.

• (Vasilescu, Langlais, Lapalme 2004) F. Vasilescu, P. Langlais, G. Lapalme "Evaluating variants of the Lesk approach for disambiguating words”, LREC 2004.

• (Yarowsky, 1993) Yarowsky, D. One sense per collocation. ARPA Workshop 1993.

Part 4:Supervised Methods of Word

Sense Disambiguation

35

69

Outline

• What is Supervised Learning?• Task Definition• Single Classifiers

– Naïve Bayesian Classifiers– Decision Lists and Trees

• Ensembles of Classifiers

70

What is Supervised Learning?

• Collect a set of examples that illustrate the various possible classifications or outcomes of an event.

• Identify patterns in the examples associated with each particular class of the event.

• Generalize those patterns into rules.• Apply the rules to classify a new event.

36

71

Learn from these examples :“when do I go to the store?”

YESNONONO4

NONONOYES3

YESNOYESNO2

NONOYESYES1

F3Ate Well?

F2Slept Well?

F1Hot Outside?

CLASSGo to Store?Day

72

Learn from these examples :“when do I go to the store?”

YESNONONO4

NONONOYES3

YESNOYESNO2

NONOYESYES1

F3Ate Well?

F2Slept Well?

F1Hot Outside?

CLASSGo to Store?Day

37

73

Outline

• What is Supervised Learning?• Task Definition• Single Classifiers

– Naïve Bayesian Classifiers– Decision Lists and Trees

• Ensembles of Classifiers

74

Task Definition

• Supervised WSD: Class of methods that induces a classifier from manually sense-tagged text using machine learning techniques.

• Resources– Sense Tagged Text– Dictionary (implicit source of sense inventory)– Syntactic Analysis (POS tagger, Chunker, Parser, …)

• Scope– Typically one target word per context– Part of speech of target word resolved– Lends itself to “targeted word” formulation

• Reduces WSD to a classification problem where a target word is assigned the most appropriate sense from a given set of possibilities based on the context in which it occurs

38

75

Sense Tagged Text

My bank/1 charges too much for an overdraft.

The University of Minnesota has an East and a West Bank/2campus right on the Mississippi River.My grandfather planted his pole in the bank/2 and got a great big catfish! The bank/2 is pretty muddy, I can’t walk there.

I went to the bank/1 to deposit my check and get a new ATM card.

Bonnie and Clyde are two really famous criminals, I think they were bank/1 robbers

76

Two Bags of Words(Co-occurrences in the “window of context”)

RIVER_BANK_BAG: a an and big campus cant catfish East got grandfather great

has his I in is Minnesota Mississippi muddy My of on planted pole pretty right River The the there University walk West

FINANCIAL_BANK_BAG: a an and are ATM Bonnie card charges check Clyde

criminals deposit famous for get I much My new overdraft really robbers the they think to too two went were

39

77

Simple Supervised Approach

Given a sentence S containing “bank”:

For each word Wi in SIf Wi is in FINANCIAL_BANK_BAG then

Sense_1 = Sense_1 + 1;If Wi is in RIVER_BANK_BAG then

Sense_2 = Sense_2 + 1;

If Sense_1 > Sense_2 then print “Financial”else if Sense_2 > Sense_1 then print “River”

else print “Can’t Decide”;

78

Supervised Methodology

• Create a sample of training data where a given target wordis manually annotated with a sense from a predetermined set of possibilities.– One tagged word per instance/lexical sample disambiguation

• Select a set of features with which to represent context.– co-occurrences, collocations, POS tags, verb-obj relations, etc...

• Convert sense-tagged training instances to feature vectors.• Apply a machine learning algorithm to induce a classifier.

– Form – structure or relation among features– Parameters – strength of feature interactions

• Convert a held out sample of test data into feature vectors.– “correct” sense tags are known but not used

• Apply classifier to test instances to assign a sense tag.

40

79

From Text to Feature Vectors

• My/pronoun grandfather/noun used/verb to/prep fish/verb along/adv the/det banks/SHORE of/prep the/detMississippi/noun River/noun. (S1)

• The/det bank/FINANCE issued/verb a/det check/noun for/prep the/det amount/noun of/prep interest/noun. (S2)

FINANCEYNYNdetverbdetS2

SHORENYNYdetprepdetadvS1

SENSE TAGinterestrivercheckfishP+2P+1P-1P-2

80

Supervised Learning Algorithms

• Once data is converted to feature vector form, any supervised learning algorithm can be used. Many have been applied to WSD with good results:– Support Vector Machines– Nearest Neighbor Classifiers– Decision Trees – Decision Lists– Naïve Bayesian Classifiers– Perceptrons– Neural Networks– Graphical Models– Log Linear Models

41

81

Outline

• What is Supervised Learning?• Task Definition• Naïve Bayesian Classifier• Decision Lists and Trees• Ensembles of Classifiers

82

Naïve Bayesian Classifier

• Naïve Bayesian Classifier well known in Machine Learning community for good performance across a range of tasks (e.g., Domingos and Pazzani, 1997)…Word Sense Disambiguation is no exception

• Assumes conditional independence among features, given the sense of a word.– The form of the model is assumed, but parameters are estimated

from training instances

• When applied to WSD, features are often “a bag of words”that come from the training data– Usually thousands of binary features that indicate if a word is

present in the context of the target word (or not)

42

83

Bayesian Inference

• Given observed features, what is most likely sense?• Estimate probability of observed features given sense • Estimate unconditional probability of sense• Unconditional probability of features is a normalizing

term, doesn’t affect sense classification

),...,3,2,1()()*|,...,3,2,1(

),...,3,2,1| ( FnFFFpSpSFnFFFp

FnFFFSp =

84

Naïve Bayesian Model

S

F2 F3 F4F1 Fn

)|(*...*)|2(*)|1()|,...,2,1( SFnpSFpSFpSFnFFP =

43

85

The Naïve Bayesian Classifier

– Given 2,000 instances of “bank”, 1,500 for bank/1 (financial sense) and 500 for bank/2 (river sense)

• P(S=1) = 1,500/2000 = .75• P(S=2) = 500/2,000 = .25

– Given “credit” occurs 200 times with bank/1 and 4 times with bank/2.• P(F1=“credit”) = 204/2000 = .102• P(F1=“credit”|S=1) = 200/1,500 = .133• P(F1=“credit”|S=2) = 4/500 = .008

– Given a test instance that has one feature “credit”• P(S=1|F1=“credit”) = .133*.75/.102 = .978• P(S=2|F1=“credit”) = .008*.25/.102 = .020

)(*)|(*...*)|1(argmax SpSFnpSFpsenseSsense∈

=

86

Comparative Results

• (Leacock, et. al. 1993) compared Naïve Bayes with a Neural Network and a Context Vector approach when disambiguating six senses of line…

• (Mooney, 1996) compared Naïve Bayes with a Neural Network, Decision Tree/List Learners, Disjunctive and Conjunctive Normal Form learners, and a perceptron when disambiguating six senses of line…

• (Pedersen, 1998) compared Naïve Bayes with Decision Tree, Rule Based Learner, Probabilistic Model, etc. when disambiguating line and 12 other words…

• …All found that Naïve Bayesian Classifier performed as well as any of the other methods!

44

87

Outline

• What is Supervised Learning?• Task Definition• Naïve Bayesian Classifiers• Decision Lists and Trees• Ensembles of Classifiers

88

Decision Lists and Trees

• Very widely used in Machine Learning. • Decision trees used very early for WSD research (e.g.,

Kelly and Stone, 1975; Black, 1988). • Represent disambiguation problem as a series of questions

(presence of feature) that reveal the sense of a word.– List decides between two senses after one positive answer– Tree allows for decision among multiple senses after a series of

answers

• Uses a smaller, more refined set of features than “bag of words” and Naïve Bayes.– More descriptive and easier to interpret.

45

89

Decision List for WSD (Yarowsky, 1994)

• Identify collocational features from sense tagged data. • Word immediately to the left or right of target :

– I have my bank/1 statement.– The river bank/2 is muddy.

• Pair of words to immediate left or right of target :– The world’s richest bank/1 is here in New York.– The river bank/2 is muddy.

• Words found within k positions to left or right of target, where k is often 10-50 :– My credit is just horrible because my bank/1 has made several

mistakes with my account and the balance is very low.

90

Building the Decision List

• Sort order of collocation tests using log of conditional probabilities.

• Words most indicative of one sense (and not the other) will be ranked highly.

))|2(

)|1((log

inCollocatioiFSpinCollocatioiFSp

Abs==

==

46

91

Computing DL score

– Given 2,000 instances of “bank”, 1,500 for bank/1 (financial sense) and 500 for bank/2 (river sense)

• P(S=1) = 1,500/2,000 = .75• P(S=2) = 500/2,000 = .25

– Given “credit” occurs 200 times with bank/1 and 4 times with bank/2.

• P(F1=“credit”) = 204/2,000 = .102• P(F1=“credit”|S=1) = 200/1,500 = .133• P(F1=“credit”|S=2) = 4/500 = .008

– From Bayes Rule…• P(S=1|F1=“credit”) = .133*.75/.102 = .978• P(S=2|F1=“credit”) = .008*.25/.102 = .020

– DL Score = abs (log (.978/.020)) = 3.89

92

Using the Decision List

• Sort DL-score, go through test instance looking for matching feature. First match reveals sense…

N/Aof the bank0.00

Bank/2 riverpole within bank1.09

Bank/2 riverbank is muddy2.20

Bank/1 financialcredit within bank3.89SenseFeatureDL-score

47

93

Using the Decision List

CREDIT?

BANK/1 FINANCIAL IS MUDDY?

POLE? BANK/2 RIVER

BANK/2 RIVER

94

Learning a Decision Tree

• Identify the feature that most “cleanly” divides the training data into the known senses.– “Cleanly” measured by information gain or gain ratio. – Create subsets of training data according to feature values.

• Find another feature that most cleanly divides a subset of the training data.

• Continue until each subset of training data is “pure” or as clean as possible.

• Well known decision tree learning algorithms include ID3 and C4.5 (Quillian, 1986, 1993)

• In Senseval-1, a modified decision list (which supported some conditional branching) was most accurate for English Lexical Sample task (Yarowsky, 2000)

48

95

Supervised WSD with Individual Classifiers

• Many supervised Machine Learning algorithms have been applied to Word Sense Disambiguation, most work reasonably well. – (Witten and Frank, 2000) is a great intro. to supervised learning.

• Features tend to differentiate among methods more than the learning algorithms.

• Good sets of features tend to include:– Co-occurrences or keywords (global)– Collocations (local)– Bigrams (local and global)– Part of speech (local)– Predicate-argument relations

• Verb-object, subject-verb,– Heads of Noun and Verb Phrases

96

Convergence of Results

• Accuracy of different systems applied to the same data tends to converge on a particular value, no one system shockingly better than another.– Senseval-1, a number of systems in range of 74-78% accuracy for

English Lexical Sample task.– Senseval-2, a number of systems in range of 61-64% accuracy for

English Lexical Sample task.– Senseval-3, a number of systems in range of 70-73% accuracy for

English Lexical Sample task…

• What to do next?

49

97

Outline

• What is Supervised Learning?• Task Definition• Naïve Bayesian Classifiers• Decision Lists and Trees• Ensembles of Classifiers

98

Ensembles of Classifiers

• Classifier error has two components (Bias and Variance)– Some algorithms (e.g., decision trees) try and build a

representation of the training data – Low Bias/High Variance– Others (e.g., Naïve Bayes) assume a parametric form and don’t

represent the training data – High Bias/Low Variance

• Combining classifiers with different bias variance characteristics can lead to improved overall accuracy

• “Bagging” a decision tree can smooth out the effect of small variations in the training data (Breiman, 1996)– Sample with replacement from the training data to learn multiple

decision trees.– Outliers in training data will tend to be obscured/eliminated.

50

99

Ensemble Considerations

• Must choose different learning algorithms with significantly different bias/variance characteristics.– Naïve Bayesian Classifier versus Decision Tree

• Must choose feature representations that yield significantly different (independent?) views of the training data.– Lexical versus syntactic features

• Must choose how to combine classifiers. – Simple Majority Voting– Averaging of probabilities across multiple classifier output– Maximum Entropy combination (e.g., Klein, et. al., 2002)

100

Ensemble Results

• (Pedersen, 2000) achieved state of art for interest and linedata using ensemble of Naïve Bayesian Classifiers.– Many Naïve Bayesian Classifiers trained on varying sized

windows of context / bags of words.– Classifiers combined by a weighted vote

• (Florian and Yarowsky, 2002) achieved state of the art for Senseval-1 and Senseval-2 data using combination of six classifiers.– Rich set of collocational and syntactic features.– Combined via linear combination of top three classifiers.

• Many Senseval-2 and Senseval-3 systems employed ensemble methods.

51

101

References• (Black, 1988) An experiment in computational discrimination of English word senses.

IBM Journal of Research and Development (32) pg. 185-194.• (Breiman, 1996) The heuristics of instability in model selection. Annals of Statistics

(24) pg. 2350-2383.• (Domingos and Pazzani, 1997) On the Optimality of the Simple Bayesian Classifier

under Zero-One Loss, Machine Learning (29) pg. 103-130.• (Domingos, 2000) A Unified Bias Variance Decomposition for Zero-One and Squared

Loss. In Proceedings of AAAI. Pg. 564-569. • (Florian an dYarowsky, 2002) Modeling Consensus: Classifier Combination for Word

Sense Disambiguation. In Proceedings of EMNLP, pp 25-32. • (Kelly and Stone, 1975). Computer Recognition of English Word Senses, North Holland

Publishing Co., Amsterdam.• (Klein, et. al., 2002) Combining Heterogeneous Classifiers for Word-Sense

Disambiguation, Proceedings of Senseval-2. pg. 87-89. • (Leacock, et. al. 1993) Corpus based statistical sense resolution. In Proceedings of the

ARPA Workshop on Human Language Technology. pg. 260-265. • (Mooney, 1996) Comparative experiments on disambiguating word senses: An

illustration of the role of bias in machine learning. Proceedings of EMNLP. pg. 82-91.

102

References

• (Pedersen, 1998) Learning Probabilistic Models of Word Sense Disambiguation. Ph.D. Dissertation. Southern Methodist University.

• (Pedersen, 2000) A simple approach to building ensembles of Naive Bayesian classifiers for word sense disambiguation. In Proceedings of NAACL.

• (Quillian, 1986). Induction of Decision Trees. Machine Learning (1). pg. 81-106.• (Quillian, 1993). C4.5 Programs for Machine Learning. San Francisco, Morgan

Kaufmann.• (Witten and Frank, 2000). Data Mining – Practical Machine Learning Tools and

Techniques with Java Implementations. Morgan-Kaufmann. San Francisco.• (Yarowsky, 1994) Decision lists for lexical ambiguity resolution: Application to accent

restoration in Spanish and French. In Proceedings of ACL. pp. 88-95.• (Yarowsky, 2000) Hierarchical decision lists for word sense disambiguation. Computers

and the Humanities, 34.

52

Part 5:Minimally Supervised Methods for Word Sense Disambiguation

104

Outline

• Task definition– What does “minimally” supervised mean?

• Bootstrapping algorithms– Co-training– Self-training– Yarowsky algorithm

• Using the Web for Word Sense Disambiguation– Web as a corpus– Web as collective mind

53

105

Task Definition

• Supervised WSD = learning sense classifiers starting with annotated data

• Minimally supervised WSD = learning sense classifiers from annotated data, with minimal human supervision

• Examples – Automatically bootstrap a corpus starting with a few human

annotated examples– Use monosemous relatives / dictionary definitions to automatically

construct sense tagged data– Rely on Web-users + active learning for corpus annotation

106

Outline




54

107

Bootstrapping WSD Classifiers

• Build sense classifiers with little training data– Expand applicability of supervised WSD

• Bootstrapping approaches– Co-training– Self-training– Yarowsky algorithm

108

Bootstrapping Recipe

• Ingredients– (Some) labeled data– (Large amounts of) unlabeled data– (One or more) basic classifiers

• Output– Classifier that improves over the basic classifiers

55

109

… plants#1 and animals …… industry plant#2 …

… building the only atomic plant …… plant growth is retarded …… a herb or flowering plant …… a nuclear power plant …… building a new vehicle plant …… the animal and plant life …… the passion-fruit plant …

Classifier 1

Classifier 2

… plant#1 growth is retarded …… a nuclear power plant#2 …

110

Co-training / Self-training

• 1. Create a pool of examples U' – choose P random examples from U

• 2. Loop for I iterations– Train Ci on L and label U'– Select G most confident examples and add to L

• maintain distribution in L– Refill U' with examples from U

• keep U' at constant size P

– A set L of labeled training examples– A set U of unlabeled examples– Classifiers Ci

56

111

• (Blum and Mitchell 1998)

• Two classifiers– independent views– [independence condition can be relaxed]

• Co-training in Natural Language Learning– Statistical parsing (Sarkar 2001)– Co-reference resolution (Ng and Cardie 2003)– Part of speech tagging (Clark, Curran and Osborne 2003)– ...

Co-training

112

Self-training

• (Nigam and Ghani 2000)

• One single classifier• Retrain on its own output• Self-training for Natural Language Learning

– Part of speech tagging (Clark, Curran and Osborne 2003)– Co-reference resolution (Ng and Cardie 2003)

• several classifiers through bagging

57

113

Parameter Setting for Co-training/Self-training

• 1. Create a pool of examples U' – choose P random examples from U

• 2. Loop for I iterations– Train Ci on L and label U'– Select G most confident examples and add to L

• maintain distribution in L– Refill U' with examples from U

• keep U' at constant size P

Pool size

Iterations

Growth size

•A major drawback of bootstrapping–“No principled method for selecting optimal values for these parameters” (Ng and Cardie 2003)

114

Experiments with Co-training / Self-training for WSD

• Training / Test data– Senseval-2 nouns (29 ambiguous nouns)– Average corpus size: 95 training examples, 48 test examples

• Raw data– British National Corpus– Average corpus size: 7,085 examples

• Co-training– Two classifiers: local and topical classifiers

• Self-training– One classifier: global classifier

• (Mihalcea 2004)

58

115

Parameter Settings • Parameter ranges

– P = {1, 100, 500, 1000, 1500, 2000, 5000}– G = {1, 10, 20, 30, 40, 50, 100, 150, 200}– I = {1, ..., 40}

• 29 nouns → 120,000 runs• Upper bound in co-training/self-training performance

– Optimised on test set– Basic classifier: 53.84%– Optimal self-training: 65.61%– Optimal co-training: 65.75%– ~25% error reduction

• Per-word parameter setting:– Co-training = 51.73%– Self-training = 52.88%

• Global parameter setting– Co-training = 55.67%– Self-training = 54.16%

• Example: lady– basic = 61.53%

– self-training = 84.61% [20/100/39]

– co-training = 82.05% [1/1000/3]

116

Yarowsky Algorithm

• (Yarowsky 1995)• Similar to co-training• Differs in the basic assumption (Abney 2002)

– “view independence” (co-training) vs. “precision independence”(Yarowsky algorithm)

• Relies on two heuristics and a decision list– One sense per collocation :

• Nearby words provide strong and consistent clues as to the sense of a target word

– One sense per discourse :• The sense of a target word is highly consistent within a single

document

59

117

Learning Algorithm• A decision list is used to classify instances of target word :

“the loss of animal and plant species through extinction …”

• Classification is based on the highest ranking rule that matches the target context

…......

→ A (living)plant species 9.02

→ A (living)fruit (within +/- k words) 9.03

→ B (factory)job (within +/- k words) 9.24

→ A (living)flower (within +/- k words) 9.31

………SenseCollocation LogL

118

Bootstrapping Algorithm

• All occurrences of the target word are identified• A small training set of seed data is tagged with word sense

Sense-B: factory

Sense-A: life

60

119


Seed set grows and residual set shrinks ….

120


Convergence: Stop when residual set stabilizes

61

121


• Iterative procedure:– Train decision list algorithm on seed set– Classify residual data with decision list – Create new seed set by identifying samples that are tagged with a

probability above a certain threshold– Retrain classifier on new seed set

• Selecting training seeds– Initial training set should accurately distinguish among possible

senses– Strategies:

• Select a single, defining seed collocation for each possible sense. Ex: “life” and “manufacturing” for target plant

• Use words from dictionary definitions• Hand-label most frequent collocates

122

Evaluation

• Test corpus: extracted from 460 million word corpus of multiple sources (news articles, transcripts, novels, etc.)

• Performance of multiple models compared with:– supervised decision lists– unsupervised learning algorithm of Schütze (1992), based on

alignment of clusters with word senses

Unsupervised Bootstrapping

96.592.296.1-Avg.

…-………

Supervised

97.9 9298.0legal/physicalmotion

96.5 9597.1vehicle/containertank

93.6 9093.9volume/outerspace

98.6 9297.7living/factoryplant

Unsupervised Schütze

Senses Word

62

123

Outline




124

The Web as a Corpus

• Use the Web as a large textual corpus– Build annotated corpora using monosemous relatives– Bootstrap annotated corpora starting with few seeds

• Similar to (Yarowsky 1995)

• Use the (semi)automatically tagged data to train WSD classifiers

63

125

Monosemous Relatives

•• IdeaIdea: determine a phrase (SP) which uniquely identifies the sense of a word (W#i)1. Determine one or more Search Phrases from a machine readable

dictionary using several heuristics 2. Search the Web using the Search Phrases from step 1.3. Replace the Search Phrases in the examples gathered at 2 with W#i.

– Output: sense annotated corpus for the word sense W#iAs a pastime, she enjoyed reading. Evaluate the interestingness of the website.

As an interest, she enjoyed reading.Evaluate the interest of the website.

126

Heuristics to Identify Monosemous Relatives

• Synonyms– Determine a monosemous synonym– remember#1 has recollect as monosemous synonym ⇒ SP=recollect

• Dictionary definitions (1)– Parse the gloss and determine the set of single phrase definitions– produce#5 has the definition “bring onto the market or release” ⇒ 2

definitions: “bring onto the market” and “release”eliminate “release” as being ambiguous ⇒ SP=bring onto the market

• Dictionary defintions (2)– Parse the gloss and determine the set of single phrase definitions– Replace the stop words with the NEAR operator– Strengthen the query: concatenate the words from the current synset using

the AND operator– produce#6 has the synset {grow, raise, farm, produce} and the definition

“cultivate by growing” ⇒SP=cultivate NEAR growing AND (grow OR raise OR farm OR produce)

64

127

• Dictionary definitions (3)– Parse the gloss and determine the set of single phrase definitions– Keep only the head phrase– Strengthen the query: concatenate the words from the current synset using

the AND operator– company#5 has the synset {party,company} and the definition “band of

people associated in some activity” ⇒SP=band of people AND (company OR party)

Heuristics to Identify Monosemous Relatives

128

Example

• Building annotated corpora for the noun interest.# Synset Definition1 {interest#1, involvement} sense of concern with and curiosity about

someone or something2 {interest#2,interestingness} the power of attracting or holding one’s interest3 {sake, interest#3} reason for wanting something done4 {interest#4} fixed charge for borrowing money; usually a

percentage of the amount borrowed5 {pastime,interest#5} a subject or pursuit that occupies one’s time and

thoughts6 {interest#6, stake} a right or legal share of something; financial

involvement with something7 {interest#7, interest group} a social group whose members control some field

of activity and who have common aims

Sense # Search phrase1 sense of concern AND (interest OR involvement)2 interestigness3 reason for wanting AND (interest OR sake)4 fixed charge AND interest

percentage of amount AND interest5 pastime6 right share AND (interest OR stake)

legal share AND (interest OR stake)financial involvement AND (interest OR stake)

7 interest group

65

129

Example

• Gather 5,404 examples• Check the first 70 examples ⇒ 67 correct; 95.7% accuracy.

1. I appreciate the genuine interest#1 which motivated you to write your message.

2. The webmaster of this site warrants neither accuracy, nor interest#2.3. He forgives us not only for our interest#3, but for his own.4. Interest#4 coverage, including rents, was 3.6x5. As an interest#5, she enjoyed gardening and taking part into church

activities.6. Voted on issues, they should have abstained because of direct and

indirect personal interests#6 in the matters of hand.7. The Adam Smith Society is a new interest#7 organized within the APA.

130

Experimental Evaluation

• Tests on 20 words – 7 nouns, 7 verbs, 3 adjectives, 3 adverbs (120 word meanings)– manually check the first 10 examples of each sense of a word

=> 91% accuracy (Mihalcea 1999)Word Polysemy

count Examples in SemCor

Total exam-ples acquired

Examples ma-nually checked

Correct examples

interest 7 139 5404 70 67 report 7 71 4196 70 63 company 9 90 6292 80 77 school 7 146 2490 59 54 produce 7 148 4982 67 60 remember 8 166 3573 67 57 write 8 285 2914 69 67 speak 4 147 4279 40 39 small 14 192 10954 107 92 clearly 4 48 4031 29 28 TOTAL (20 words)

120 2582 80741 1080 978

66

131

Web-based Bootstrapping

• Similar to Yarowsky algorithm• Relies on data gathered from the Web1. Create a set of seeds (phrases) consisting of:

– Sense tagged examples in SemCor– Sense tagged examples from WordNet– Additional sense tagged examples, if available

• Phrase?– At least two open class words; – Words involved in a semantic relation (e.g. noun phrase, verb-object,

verb-subject, etc.)2. Search the Web using queries formed with the seed expressions found

at Step 1– Add to the generated corpus of maximum of N text passages– Results competitive with manually tagged corpora (Mihalcea 2002)

132

The Web as Collective Mind

• Two different views of the Web:– collection of Web pages– very large group of Web users

• Millions of Web users can contribute their knowledge to a data repository

• Open Mind Word Expert (Chklovski and Mihalcea, 2002)

• Fast growing rate: – Started in April 2002– Currently more than 100,000 examples of noun senses in several

languages

67

133

OMWEonline

http://teach-computers.org

134

Open Mind Word Expert: Quantity and Quality• Data

– A mix of different corpora: Treebank, Open Mind Common Sense, Los Angeles Times, British National Corpus

• Word senses– Based on WordNet definitions

• Active learning to select the most informative examples for learning– Use two classifiers trained on existing annotated data– Select items where the two classifiers disagree for human annotation

• Quality: – Two tags per item– One tag per item per contributor

• Evaluations:– Agreement rates of about 65% - comparable to the agreements

rates obtained when collecting data for Senseval-2 with trained lexicographers

– Replicability: tests on 1,600 examples of “interest” led to 90%+ replicability

68

135

References• (Abney 2002) Abney, S. Bootstrapping. Proceedings of ACL 2002.• (Blum and Mitchell 1998) Blum, A. and Mitchell, T. Combining labeled and unlabeled

data with co-training. Proceedings of COLT 1998.• (Chklovski and Mihalcea 2002) Chklovski, T. and Mihalcea, R. Building a sense tagged

corpus with Open Mind Word Expert. Proceedings of ACL 2002 workshop on WSD.• (Clark, Curran and Osborne 2003) Clark, S. and Curran, J.R. and Osborne, M.

Bootstrapping POS taggers using unlabelled data. Proceedings of CoNLL 2003.• (Mihalcea 1999) Mihalcea, R. An automatic method for generating sense tagged

corpora. Proceedings of AAAI 1999.• (Mihalcea 2002) Mihalcea, R. Bootstrapping large sense tagged corpora. Proceedings

of LREC 2002.• (Mihalcea 2004) Mihalcea, R. Co-training and Self-training for Word Sense

Disambiguation. Proceedings of CoNLL 2004.• (Ng and Cardie 2003) Ng, V. and Cardie, C. Weakly supervised natural language

learning without redundant views. Proceedings of HLT-NAACL 2003.• (Nigam and Ghani 2000) Nigam, K. and Ghani, R. Analyzing the effectiveness and

applicability of co-training. Proceedings of CIKM 2000.• (Sarkar 2001) Sarkar, A. Applying cotraining methods to statistical parsing. Proceedings

of NAACL 2001.• (Yarowsky 1995) Yarowsky, D. Unsupervised word sense disambiguation rivaling

supervised methods. Proceedings of ACL 1995.

Part 6: Unsupervised Methods of Word

Sense Disambiguation

69

137

Outline

• What is Unsupervised Learning?• Task Definition• Agglomerative Clustering• LSI/LSA• Sense Discrimination Using Parallel Texts

138

What is Unsupervised Learning?

• Unsupervised learning identifies patterns in a large sample of data, without the benefit of any manually labeled examples or external knowledge sources

• These patterns are used to divide the data into clusters, where each member of a cluster has more in common with the other members of its own cluster than any other

• Note! If you remove manual labels from supervised data and cluster, you may not discover the same classes as in supervised learning– Supervised Classification identifies features that trigger a sense tag– Unsupervised Clustering finds similarity between contexts

70

139

Cluster this Data!Facts about my day…

YESNONO4

NONONO3

YESNOYES2

NONOYES1

F3Ate Well?

F2Slept Well?

F1Hot Outside?Day

140

Cluster this Data!Facts about my day…

YESNONO4

NONONO3

YESNOYES2

NONOYES1

F3Ate Well?

F2Slept Well?

F1Hot Outside?Day

71

141

Cluster this Data!

YESNONO4

NONONO3

YESNOYES2

NONOYES1

F3Ate Well?

F2Slept Well?

F1Hot Outside?Day

142

Outline


72

143

Task Definition

• Unsupervised Word Sense Discrimination: A class of methods that cluster words based on similarity of context

• Strong Contextual Hypothesis – (Miller and Charles, 1991): Words with similar meanings tend to

occur in similar contexts– (Firth, 1957): “You shall know a word by the company it keeps.”

• …words that keep the same company tend to have similar meanings

• Only use the information available in raw text, do not use outside knowledge sources or manual annotations

• No knowledge of existing sense inventories, so clusters are not labeled with senses

144

Task Definition

• Resources:– Large Corpora

• Scope: – Typically one targeted word per context to be discriminated– Equivalently, measure similarity among contexts– Features may be identified in separate “training” data, or in the

data to be clustered– Does not assign senses or labels to clusters

• Word Sense Discrimination reduces to the problem of finding the targeted words that occur in the most similar contexts and placing them in a cluster

73

145

Outline


146

Agglomerative Clustering

• Create a similarity matrix of instances to be discriminated– Results in a symmetric “instance by instance” matrix, where each

cell contains the similarity score between a pair of instances– Typically a first order representation, where similarity is based on

the features observed in the pair of instances

• Apply Agglomerative Clustering algorithm to matrix– To start, each instance is its own cluster– Form a cluster from the most similar pair of instances– Repeat until the desired number of clusters is obtained

• Advantages : high quality clustering • Disadvantages – computationally expensive, must carry

out exhaustive pair wise comparisons)()(

YXYX

∪∩

74

147

Measuring Similarity

• Integer Values

– Matching Coefficient

– Jaccard Coefficient

– Dice Coefficient

• Real Values

– Cosine

YX ∩

YXYX

∪∩

YXYX

+∩×2

YX

YXrr

⋅

148

Instances to be Clustered

NNNYdetverbadjdetS3YNYNdetprepdetS2

NNNNnounnounnoundetS4

NYNYdetprepdetadvS1

interestrivercheckfishP+2P+1P-1P-2

1

24

S3

124S302S4

03S223S1

S4S2S1

75

149

Average Link Clusteringaka McQuitty’s Similarity Analysis

1

2

4

S3

124S3

02S4

03S2

23S1

S4S2S1

0S4

0S2

S1S3

S4S2S1S3

5.22

23=

+

5.22

23=

+

5.12

12=

+

5.12

12=

+

S4

S1S3S2

S4S1S3S2

5.12

5.15.1=

+

5.12

5.15.1=

+

150

Evaluation of Unsupervised Methods

• If Sense tagged text is available, can be used for evaluation– But don’t use sense tags for clustering or feature selection!

• Assume that sense tags represent “true” clusters, and compare these to discovered clusters– Find mapping of clusters to senses that attains maximum accuracy

• Pseudo words are especially useful, since it is hard to find data that is discriminated– Pick two words or names from a corpus, and conflate them into

one name. Then see how well you can discriminate.– http://www.d.umn.edu/~kulka020/kanaghaName.html

• Baseline Algorithm– group all instances into one cluster, this will reach “accuracy” equal to majority classifier

76

151

Baseline Performance

170553580Totals

170553580C3

0000C2

0000C1

TotalsS3S2S1

170803555Totals

170803555C3

0000C2

0000C1

TotalsS1S2S3

(0+0+55)/170 = .32 (0+0+80)/170 = .47if C3 is S3 if C3 is S1

152

Evaluation

• Suppose that C1 is labeled S1, C2 as S2, and C3 as S3• Accuracy = (10 + 0 + 10) / 170 = 12% • Diagonal shows how many members of the cluster actually belong to

the sense given on the column • Can the “columns” be rearranged to improve the overall accuracy?

– Optimally assign clusters to senses

170553580Totals

6510550C3

6040020C2

4553010C1

TotalsS3S2S1

77

153

Evaluation

• The assignment of C1 to S2, C2 to S3, and C3 to S1 results in 120/170 = 71%

• Find the ordering of the columns in the matrix that maximizes the sum of the diagonal.

• This is an instance of the Assignment Problem from Operations Research, or finding the Maximal Matching of a Bipartite Graph from Graph Theory.

170805535Totals

6550105C3

6020400C2

4510530C1

TotalsS1S3S2

154

Agglomerative Approach

• (Pedersen and Bruce, 1997) explore discrimination with a small number (approx 30) of features near target word.– Morphological form of target word (1)– Part of Speech two words to left and right of target word (4)– Co-occurrences (3) most frequent content words in context– Unrestricted collocations (19) most frequent words located one

position to left or right of target, OR– Content collocations (19) most frequent content words located one

position to left or right of target • Features identified in the instances be clustered• Similarity measured by matching coefficient• Clustered with McQuitty’s Similarity Analysis, Ward’s

Method, and the EM Algorithm– Found that McQuitty’s method was the most accurate

78

155

Experimental Evaluation

• Adjectives– Chief, 86% majority (1048)– Common, 84% majority (1060)– Last, 94% majority (3004)– Public, 68% majority (715)

• Nouns– Bill, 68% majority (1341)– Concern, 64% majority (1235)– Drug, 57% majority (1127)– Interest, 59% majority (2113)– Line, 37% majority (1149)

• Verbs– Agree, 74% majority (1109)– Close, 77% majority (1354)– Help, 78% majority (1267)– Include, 91% majority (1526)

• Adjectives– Chief, 86% – Common, 80% – Last, 79% – Public, 63%

• Nouns– Bill, 75%– Concern, 68%– Drug, 65%– Interest, 65%– Line, 42%

• Verbs– Agree, 69%– Close, 72%– Help, 70%– Include, 77%

156

Analysis

• Unsupervised methods may not discover clusters equivalent to the classes learned in supervised learning

• Evaluation based on assuming that sense tags represent the “true” cluster are likely a bit harsh. Alternatives?– Humans could look at the members of each cluster and determine

the nature of the relationship or meaning that they all share– Use the contents of the cluster to generate a descriptive label that

could be inspected by a human

• First order feature sets may be problematic with smaller amounts of data since these features must occur exactly in the test instances in order to be “matched”

79

157

Outline


158

Latent Semantic Indexing/Analysis

• Adapted by (Schütze, 1998) to word sense discrimination• Represent training data as word co-occurrence matrix• Reduce the dimensionality of the co-occurrence matrix via

Singular Value Decomposition (SVD)– Significant dimensions are associated with concepts

• Represent the instances of a target word to be clustered by taking the average of all the vectors associated with all the words in that context– Context represented by an averaged vector

• Measure the similarity amongst instances via cosine and record in similarity matrix, or cluster the vectors directly

80

159

Co-occurrence matrix

4

2

0

0

0

3

0

1

box

0

1

2

2

1

2

0

0

memory

0

0

0

1

0

0

2

0

organ

02032000debt

01031002linux

01032000sales

30220300lab

10200120petri

01002001disk

10200030body

00031002pc

plasmagraphicstissuedataibmcellsbloodapple

160

Singular Value DecompositionA=UDV’

81

161

U

-.52

.39

-.48

.02

.09

.41

-.09.40

-.30

.08

.31

.43

-.26

-.39

-.6.20

.00

-.00

-.00

-.02

-.01

.00

-.02-.00

-.07-.3.14-.49-.07.30.25.56

-.01.08.05-.01.24-.08.11.46

.08.03-.04.72.09-.31-.01.37

-.07.01-.21-.31-.34-.45-.68.29

.00.05.83.17-.02.25-.45.08

.03.20-.22.31-.60.39.13.35

-.01-.04-.44.08.44.59-.49.05-.02.63.02-.09.52-.2.09.35

162

D

0.00

0.00

0.00

0.661.26

2.30

2.52

3.25

3.99

6.369.19

82

163

V

-.20.22-.07-.10-.87-.07-.06.17.19-.26.04.03

.17

-.32.02

.13

-.26

-.17.06-.04

.86

.50-.58.12.09-.18-.27-.18-.12-.47.11

-.03.12.31-.32-.04.64-.45-.14-.23.28

.07-.23-.62-.59.05.02-.12.15.11.25-.71-.31-.04.08.29-.05.05.20-.51.09

-.03.12.31-.01.02-.45-.32.50.27.49

-.02.08.21-.06.08-.09.52-.45-.01.63

.03-.12-.31.71-.13.39-.12.12.15.37

.07.58-.41.15.17-.30-.32-.27-.39.11

.44.25.03-.02.26.23.39.57-.37.04

.03-.12-.31-.05-.05.04.28-.04.08.21

164

Co-occurrence matrix after SVD

1.1

1.0

.98

1.7

.86

.72

.85

.77

memory

.00

.00

.17

1.2

.77

.00

.84

.00

organ

.001.5.003.22.1.00.001.2debt

.131.1.032.71.7.16.00.96linux

.41.85.352.21.3.39.15.73sales

2.3.182.51.7.352.01.7.21lab

1.4.001.5.49.001.21.1.00germ

.00.91.002.11.3.01.00.76disk

1.5.001.6.33.001.31.2.00body

.09.86.012.01.3.11.00.73pc

plasmagraphicstissuedataibmcellsbloodapple

83

165

Effect of SVD

• SVD reduces a matrix to a given number of dimensions This may convert a word level space into a semantic or conceptual space– If “dog” and “collie” and “wolf” are dimensions/columns in the

word co-occurrence matrix, after SVD they may be a single dimension that represents “canines”

• The dimensions are the principle components that may (hopefully) represent the meaning of concepts

• SVD has effect of smoothing a very sparse matrix, so that there are very few 0 valued cells

166

Context Representation

• Represent each instance of the target word to be clustered by averaging the word vectors associated with its context– This creates a “second order” representation of the context

• The context is represented not only by the words that occur therein, but also the words that occur with the words in the context elsewhere in the training corpus

84

167

Second Order Context Representation

• These two contexts share no words in common, yet they are similar! disk and linux both occur with “Apple”, “IBM”, “data”, “graphics”, and “memory”

• The two contexts are similar because they share many second order co-occurrences

1.0.72

memory

.00

.00

organ

.131.1.032.71.7.16.00.96linux

.00.91.002.11.3.01.00.76disk

Plasmagraphicstissuedataibmcellsbloodapple

• I got a new disk today!• What do you think of linux?

168

Second Order Context Representation

• The bank of the Mississippi River was washed away.

85

169

First vs. Second Order Representations

• Comparison made by (Purandare and Pedersen, 2004) • Build word co-occurrence matrix using log-likelihood ratio

– Reduce via SVD– Cluster in vector or similarity space– Evaluate relative to manually created sense tags

• Experiments conducted with Senseval-2 data – 24 words, 50-200 training and test examples– Second order representation resulted in significantly better

performance than first order, probably due to modest size of data.

• Experiments conducted with line, hard, serve– 4000-5000 instances, divided into 60-40 training-test split– First order representation resulted in better performance than

second order, probably due to larger amount of data

170

Analysis

• Agglomerative methods based on direct (first order) features require large amounts of data in order to obtain a reliable set of features

• Large amounts of data are problematic for agglomerative clustering (due to exhaustive comparisons)

• Second order representations allow you to make due with smaller amounts of data, and still get a rich (non-sparse) representation of context

• http://senseclusters.sourceforge.net is a complete system for performing unsupervised discrimination using first or second order context vectors in similarity or vector space, and includes support for SVD, clustering and evaluation

86

171

Outline


172

Sense Discrimination Using Parallel Texts

• There is controversy as to what exactly is a “word sense”(e.g., Kilgarriff, 1997)

• It is sometimes unclear how fine grained sense distinctions need to be to be useful in practice.

• Parallel text may present a solution to both problems!– Text in one language and its translation into another

• Resnik and Yarowsky (1997) suggest that word sense disambiguation concern itself with sense distinctions that manifest themselves across languages.– A “bill” in English may be a “pico” (bird jaw) in or a “cuenta”

(invoice) in Spanish.

87

173

Parallel Text

• Parallel Text can be found on the Web and there are several large corpora available (e.g., UN Parallel Text, Canadian Hansards)

• Manual annotation of sense tags is not required! However, text must be word aligned (translations identified between the two languages). – http://www.cs.unt.edu/~rada/wpt/

Workshop on Parallel Text, NAACL 2003

• Given word aligned parallel text, sense distinctions can be discovered. (e.g., Li and and Li, 2002, Diab, 2002)

174

References• (Diab, 2002) Diab, Mona and Philip Resnik, An Unsupervised Method for Word Sense

Tagging using Parallel Corpora, Proceedings of ACL, 2002. • (Firth, 1957) A Synopsis of Linguistic Theory 1930-1955. In Studies in Linguistic

Analysis, Oxford University Press, Oxford. • (Kilgarriff, 1997) “I don’t believe in word senses”, Computers and the Humanities (31)

pp. 91-113.• (Li and Li, 2002) Word Translation Disambiguation Using Bilingual Bootstrapping.

Proceedings of ACL. Pp. 343-351.• (McQuitty, 1966) Similarity Analysis by Reciprocal Pairs for Discrete and Continuous

Data. Educational and Psychological Measurement (26) pp. 825-831. • (Miller and Charles, 1991) Contextual correlates of semantic similarity. Language and

Cognitive Processes, 6 (1) pp. 1 - 28.• (Pedersen and Bruce, 1997) Distinguishing Word Sense in Untagged Text. In

Proceedings of EMNLP2. pp 197-207.• (Purandare and Pedersen, 2004) Word Sense Discrimination by Clustering Contexts in

Vector and Similarity Spaces. Proceedings of the Conference on Natural Language and Learning. pp. 41-48.

• (Resnik and Yarowsky, 1997) A Perspective on Word Sense Disambiguation Methods and their Evaluation. The ACL-SIGLEX Workshop Tagging Text with Lexical Semantics. pp. 79-86.

• (Schutze, 1998) Automatic Word Sense Discrimination. Computational Linguistics, 24 (1) pp. 97-123.

88

Part 7:How to Get Started in

Word Sense Disambiguation Research

176

Outline

• Where to get the required ingredients?– Machine Readable Dictionaries– Machine Learning Algorithms– Sense Annotated Data– Raw Data

• Where to get WSD software?• How to get your algorithms tested?

– Senseval

89

177

Machine Readable Dictionaries

• Machine Readable format (MRD)– Oxford English Dictionary– Collins– Longman Dictionary of Ordinary Contemporary English (LDOCE)

• Thesauri – add synonymy information– Roget Thesaurus http://www.thesaurus.com

• Semantic networks – add more semantic relations– WordNet http://www.cogsci.princeton.edu/~wn/

• Dictionary files, source code– EuroWordNet http://www.illc.uva.nl/EuroWordNet/

• Seven European languages

178

Machine Learning Algorithms

• Many implementations available online• Weka: Java package of many learning algorithms

– http://www.cs.waikato.ac.nz/ml/weka/– Includes decision trees, decision lists, neural networks, naïve

bayes, instance based learning, etc.

• C4.5: C implementation of decision trees– http://www.cse.unsw.edu.au/~quinlan/

• Timbl: Fast optimized implementation of instance based learning algorithms– http://ilk.kub.nl/software.html

• SVM Light: efficient implementation of Support Vector Machines– http://svmlight.joachims.org

90

179

Sense Tagged Data

• A lot of annotated data available through Senseval– http://www.senseval.org

• Data for lexical sample– English (with respect to Hector, WordNet, Wordsmyth)– Basque, Catalan, Chinese, Czech, Romanian, Spanish, etc.– Data produced within Open Mind Word Expert project

http://teach-computers.org• Data for all words

– English, Italian, Czech (Senseval-2 and Senseval-3)– SemCor (200,000 running words)

http://www.cs.unt.edu/~rada/downloads.html

• Pointers to additional data available from– http://www.senseval.org/data.html

180

Sense Tagged Data – Lexical Sample<instance id="art.40008" docsrc="bnc_ANF_855"><answer instance="art.40008" senseid="art%1:06:00::"/><context>The evening ended in a brawl between the different factions in Cubism, but it brought a

moment of splendour into the blackouts and bombings of war. [/p] [p] Yet Modigliani was too much a part of the life of Montparnasse, too involved with the

individuals leading the " new art " , to remain completely aloof.In 1914 he had met Hans Arp, the French painter who was to become prominent in the

new Dada movement, at the artists' canteen in the Avenue du Maine.Two years later Arp was living in Zurich, a member of a group of talented emigrant

artists who had left their own countries because of the war.Through casual meetings at cafes, the artists drew together to form a movement in

protest against the waste of war, against nationalism and against everything pompous, conventional or boring in the <head>art</head>of the Western world.

</context></instance>

91

181

Sense Tagged Data – SemCor <p pnum=1><s snum=1><wf cmd=ignore pos=DT>The</wf><wf cmd=done rdf=group pos=NNP lemma=group wnsn=1 lexsn=1:03:00::

pn=group>Fulton_County_Grand_Jury</wf><wf cmd=done pos=VB lemma=say wnsn=1 lexsn=2:32:00::>said</wf><wf cmd=done pos=NN lemma=friday wnsn=1 lexsn=1:28:00::>Friday</wf><wf cmd=ignore pos=DT>an</wf><wf cmd=done pos=NN lemma=investigation wnsn=1

lexsn=1:09:00::>investigation</wf><wf cmd=ignore pos=IN>of</wf><wf cmd=done pos=NN lemma=atlanta wnsn=1 lexsn=1:15:00::>Atlanta</wf><wf cmd=ignore pos=POS>'s</wf><wf cmd=done pos=JJ lemma=recent wnsn=2 lexsn=5:00:00:past:00>recent</wf><wf cmd=done pos=NN lemma=primary_election wnsn=1

lexsn=1:04:00::>primary_election</wf><wf cmd=done pos=VB lemma=produce wnsn=4 lexsn=2:39:01::>produced</wf>…

182

Raw Data

• For use with– Bootstrapping algorithms– Word sense discrimination algorithms

• British National Corpus – 100 million words covering a variety of genres, styles– http://www.natcorp.ox.ac.uk/

• TREC (Text Retrieval Conference) data– Los Angeles Times, Wall Street Journal, and more– 5 gigabytes of text– http://trec.nist.gov/

• The Web

92

183

Outline



– Senseval

184

WSD Software – Lexical Sample

• Duluth Senseval-2 systems– A decision trees system that participated in Senseval-2– http://www.d.umn.edu/~tpederse/senseval2.html

• SenseTools– For easy implementation of supervised WSD– Transforms Senseval-formatted data into the files required by Weka– http://www.d.umn.edu/~tpederse/sensetools.html

• SenseRelate– Identifies the sense of a word based on the semantic relation with its

neighbors– http://www.d.umn.edu/~tpederse/senserelate.html– Uses WordNet::Similarity – metrics of similarity on WordNet

• http://www.d.umn.edu/~tpederse/similarity.html

93

185

WSD Software – All Words

• SenseLearner– A minimally supervised approach for all open class words – Extension of a system participating in Senseval-3– http://lit.csci.unt.edu/~senselearner

• SenseRelate::AllWords– Identifies the sense of a word based on the semantic relation with

its neighbors– http://www.d.umn.edu/~tpederse/senserelate.html

186

WSD Software – Unsupervised

• SenseClusters– Finds sense clusters in a corpus of raw texts– http://www.d.umn.edu/~tpederse/senseclusters.html

94

187

Outline



– Senseval

188

Senseval

• Evaluation of WSD systems http://www.senseval.org• Senseval 1: 1999 – about 10 teams• Senseval 2: 2001 – about 30 teams• Senseval 3: 2004 – about 55 teams• Senseval 4: 2007(?)

• Provides sense annotated data for many languages, for several tasks– Languages: English, Romanian, Chinese, Basque, Spanish, etc.– Tasks: Lexical Sample, All words, etc.

• Provides evaluation software• Provides results of other participating systems

95

189

Senseval

Senseval 1 Senseval 2 Senseval 3 0

20

40

60

80

100

120

140

160

Senseval Evaluations

TasksTeamsSystems

Part 8:Conclusions

96

191

Outline

• The Web and WSD• Multilingual WSD• The Next Five Years (2005-2010)• Concluding Remarks

192

The Web and WSD

• The Web has become a source of data for NLP in general, and word sense disambiguation is no exception.

• Can find hundreds/thousands(?) of instances of a particular target word just by searching.

• Search Engines :– Alta Vista – allows scraping, at a modest rate. Insert 5 second

delays on your queries to Alta-Vista so as to not overwhelm the system. No API provided, but Perl::LWP works nicely.

http://search.cpan.org/dist/libwww-perl/– Google – does not allow scraping, but provides an API to access

search engine. However, the API limits you to 1,000 queries per day.

http://www.google.com/apis/

97

193

The Web and WSD

• The Web can search as a good source of information for selecting or verifying collocations and other kinds of association.– “strong tea” : 13,000 hits– “powerful tea” : 428 hits– “sparkling tea” : 376 hits

194

The Web and WSD

• You can find sets of related words from the Web. – http://labs.google.com/sets– Give Google Sets two or three words, it will return a set of words it

believes are related– Could be the basis of extending features sets for WSD, since many

times the words are related in meaning• Google Sets Input: bank, credit• Google Sets Output: bank, credit, stock, full, investment, invoicing,

overheads, cash low, administration, produce service, grants, overdue notices

– Great source of info about names or current events• Google Sets Input: Nixon, Carter• Google Sets Output: Carter, Nixon, Reagan, Ford, Bush, Eisenhower,

Kennedy, Johnson

98

195

A Natural Problem for the Web and WSD

• Organize Search Results by concepts, not just names.– Separate the Irish Republican Army (IRA) from the Individual

Retirement Account (IRA).

• http://clusty.com is an example of a web site that attempts to cluster content. – Finds a set of pages, and labels them with some descriptive term. – Very similar to problem in word sense discrimination, where

cluster is not associated with a known sense.

196

The Web and WSD, not all good news

• Lots and lots of junk to filter through. Lots of misleading and malicious content on web pages.

• Counts as reported by search engines for hits are approximations and vary sometime from query to query. Over time they will change, so it’s very hard to reproduce experimental results over time.

• Search engines could close down API, prohibit scraping, etc. – there are no promises made.

• Can be slow to get data from the Web.

99

197

Multilingual WSD

• Parallel text is a potential meeting ground between raw untagged text (like unsupervised methods use) and sense tagged text (like the supervised methods need)

• A source language word that is translated into various different target language forms may be polysemous in the source language

198

A Clever Way to Sense Tag

• Expertise of native speakers can be used to create sense tagged text, without having to refer to dictionaries!

• Have a bilingual native speaker pick the proper translation for a word in a given context.

http://www.teach-computers.org/word-expert/english-hindi/http://www.teach-computers.org/word-expert/english-french/

• This is a much more intuitive way to sense tag text, and depends only on the native speakers expertise, not a set of senses as found in a particular dictionary.

100

199

The Next Five Years

• Applications, applications, applications, and applications. Where are the applications? WSD needs to demonstrate an impact on applications in the next five years.

• Word Sense Disambiguation will be deployed in an increasing number of applications over the next five years.– However, not in Machine Translation. Too difficult to integrate

WSD into current statistical systems, and this won’t change soon.– Most likely applications include web search tools and email

organizers and search tools (like gmail).• If you are writing papers, “bake off” evaluations will meet

with more rejection that acceptance• If you have a potential application for Word Sense

Disambiguation in any of its forms, tell us!! Please!

200

Concluding Remarks

• Word Sense Disambiguation has something for everyone!– Statistical Methods – Knowledge Based systems– Supervised Machine Learning– Unsupervised Learning– Semi-Supervised– Bootstrapping and Co-training– Human Annotation of Data

• The impact of high quality WSD will be huge. NLP consumers have become accustomed to systems that only make coarse grained distinctions between concepts, or who might not make any at all.

• Real Understanding? Real AI?

word sense disambiguation - semantic scholar 18th european summer school in logic, language and...

Documents