Anaphora Resolution Sobha Lalitha Devi AU-KBC Research Centre MIT Campus of Anna University Chennai-44 [email protected]

05/24/2011 Summer School, IIIT HYderabad 1

Anaphora Resolution

Sobha Lalitha DeviAU-KBC Research CentreMIT Campus of Anna [email protected]


Contents

Introduction to Anaphora and Anaphora Resolution

Types of Anaphora Process of Anaphora Resolution Tools Applications References


What is Cohesion

COHESION is the internal continuity or network of points of continuity within a text.

Text is not just a string of sentences. It is not simply a large grammatical unit

“something of the same kind as a sentence, but differing from it in size- a sort of super-sentence· A semantic unit”

Halliday & Hassan


Cohesive Relationships

Cohesive relationships between words and sentences have certain definable qualities that allow us to recognize the super sentence

Nature of cohesive relation Type of cohesion

Relatedness of form Substitution and ellipsis; Lexical collocation

Relatedness of reference Reference, Lexical reiteration

Semantic connection Conjunction


Relatedness of Form

Substitution: "Nice teapots! I'll take one.“

Ellipsis: "Turn on. Tune in. Drop out." ['you' is elided]

Collocation: "John went to the bank. He wanted to swim in the river." ['river' disambiguates 'bank']


Relatedness of Reference

Exophora: (extra linguistic feature: deitic markers , this that)

“what is this?”Anaphora:

"I used to have the key. But I lost it.”Cataphora:

"It is your turn, John”Reiteration: "He speaks only to the Huxleys; the Huxleys speak only

to the Darwins; and the Darwins speak only to God.“


Semantic Connection

Conjunction: "You tell me that you've got ev'rything you want, and your

bird can sing, but you don't get me, you don't get me!

You say you've seen seven wonders, and your bird is green, but you can't see me, you can't see me!

When your prized possessions, start to tear you down, then look in my direction, I'll be round, I'll be round."

[Beatles -- Lee Campbell]


Comparison

Halliday & Hassan also classify comparison as a form of cohesion

"She's more fun than a barrel of monkeys!"

"He's as tall as a six foot four inch tree."


CONTEXT DEPENDENCE

The interpretation of most expressions depends on the context in which they are used

Developing methods for interpreting context dependent expressions useful in many applications

We focus here on dependence of nominal expressions on context introduced LINGUISTICALLY,

for which we will use the term ANAPHORA


Introduction

What is Anaphora AntecedentAnaphora Resolution

1. Sabeer Bhatia arrived at Los Angeles International Airport at 6 p.m. on September 23, 1998. His flight from Bangalore had taken 22hrs and he was starving.

[RD, NOV 2000]


Etymology of Anaphora

ANA- Back, Upstream, Back upstream

Phora- Act of Carrying

Anaphora - Act of Carrying Back


What is Anaphora

Anaphora, in discourse, is a device for making an abbreviated reference (containing fewer bits of disambiguating information, rather than being lexically or phonetically shorter) to some entity (or entities) in the expectation that the receiver of the discourse will be able to disabbreviate the reference and, thereby, determine the identity of the entity.

(Hirst 1981)


Cataphora

When “anphora” precedes the antecedent

Because she was going to the departmental store, Mary was asked to pick up the vegetables.


Relevance from the Linguistics point of view

Binding Theory is one of the major results of the principles and parameters approach developed in Chomsky (1981) and is one of the mainstays of generative linguistics.

The Binding Theory deals with the relations between nominal expressions and possible antecedents.

It attempts to provide a structural account of the complementarity of distribution between pronouns, reflexives and R-expressions.


Dichotomy Between Linguistic and NLP

The Binding Theory (and its various formulations) deals only with intra-sentential anaphora,

A very small subset of the anaphoric phenomenon that practical NLP systems are interested in resolving.

A much larger set of anaphoric phenomenon is the resolution of pronouns inter-sententially.

This problem is dealt with by Discourse Representation Theory and more specifically by Centering Theory (Grosz et al., 1995)..


Types of Anaphors

The Prime Minister is yet to arrive and he is expected at the central hall at any time. [The Times of India, Feb 2001]

This book is about Anaphora Resolution. The book is designed to help beginners in the field and its author hopes that it will be useful.

VP Anaphor John screamed, as did Mary .


Pronominal anaphora Vajpayee hits back forcefully when he told the

opposition today “sometimes we fall prey to the media and sometimes you do. [Indian Express 2001]

Possessive Priyanka eats only chicken sandwiches

before going to take any exam; nothing else goes down her gullet that day.[Indian Express, 13 March 2001]


Reflexive Pronoun

Finally ,Danian heaved himself up and lay on a waiting stretcher.

Demonstrative PronounJohn had lots of packing to do before he shifted his

house. This was something he never liked….

Relative PronounStumper Sameer Dige, who made his test debut, failed

to show fast reflexives when it mattered.


Non Anaphoric Usage of Pronouns.

Pleonastic ItCognativea. It is believed that…..b. It appears that…..Modal Adjectivesc. It is dangerous……d. It is important…..Temporale. It is five o’clock f. It is winterWeather verbsg. It is rainingf. It is snowingDistanceh. How far it is to Chennai?


Non-anaphoric uses of pronounsHe that plants thorns must never expect to

gather roses.He who dares wins.

DeicticHe seems remarkably bright for a child of his

age.


Noun Phrase Anaphora

Definite descriptions and Proper names

Roy Kaene has warned Manchester United he may snub their pay deal. United’s skipper is even hinting that unless the future Old Trafford Package meets his demands, he could quit the club in June 2000. Irishman Keane, 27, still has 17 months to run on his current 23,000 pound a week contract and wants to commit himself to United for life. Alex Ferguson’s No 1 player confirmed: If it’s not the contract I want, I won’t sign”.


Coreference

Computational Linguists from many different countries attended the tutorial. The participants found it hard to cope with the speed of the presentation, nevertheless they manages to take extensive notes.


Coreference

Sophia Loren says she will always be grateful to Bono. The actress revealed that the U2 singer helped her calm down when she became scared by a thunderstorm while travelling by a plane.

She=> Sophia LorenThe actress=> Sophia LorenThe U2 Singer=> BonoHer=>Sophia LorenShe=>Sophia Loren


Coreference chain

Sophia Loren says she will always be grateful to Bono. The actress revealed that the U2 singer helped her calm down when she became scared by a thunderstorm while travelling by a plane.

Coreference chains {Sophia Loren, she, the actress, her, she} {Bono, the U2 singer}


Chains of object mentions in text

Toni Johnson pulls a tape measure across the front of what was once a stately Victorian home. A deep trench now runs along its north wall, exposed when the house lurched two feet off its foundation during last week's earthquake.

Once inside, she spends nearly four hours measuring and diagramming each room in the 80-year-old house, gathering enough information to estimate what it would cost to rebuild it.

While she works inside, a tenant returns with several friends to collect furniture and clothing.

One of the friends sweeps broken dishes and shattered glass from a countertop and starts to pack what can be salvaged from the kitchen.

(WSJ section of Penn Treebank corpus)


What is Anaphora Resolution

The Process of finding the antecedent for an Anaphor is Anaphora resolution

Anaphor-The reference that point to the previous item.

Antecedent-The entity to which the anaphor refers


RESEARCH ON ANAPHORA RESOLUTION: A QUICK SUMMARY

1970-1995Primarily theoreticalEmphasis: commonsense knowledge, salienceException: Hobbs 1977, Shalom Lappin

1995-2005First annotated corpora to be used to develop, evaluate and compare systems (MUC, Geand Charniak, ACE)First robust systems Heuristic-based: MitkovML: Vieira & Poesio 1998, 2000; Soon et al 2001, Ng and Cardie2002Emphasis: surface featuresExceptions: Poesio & Vieira, Harabagiu, Markert

2005-presentMore sophisticated ML techniques (global models, kernels)Richer features –especially semantic informationFirst tools


Application of Anaphora Resolution

Tasks that require determining the coherence of (segments of) textSegmentation

Post-hoc coherence check in summarization (Steinberger et al, 2007)

Tasks that require identifying the most important information ina text

Sentence selection in summarization (Steinberger et al 2005, 2007)Indexing

Information extraction: recognize which expressions refer to objects in the domain

Relation extraction from biomedical text (Sanchez-Grailletand Poesio, 2006, 2007)

Multimodal interfaces: recognize which objects in the visual scene are being referred to


Different Approaches In Anaphora Resolution

Rule Based Statistical Based Machine Learning Based


Rule Based

Hobbs system


Hard Constraints on Coreference

Number agreement Person and case Gender Agreement Syntactic Agreement Selectional Restrictions


Number agreement

John and Mary loaned Sue a cup of coffee. Little did they know the magnitude of her addiction.

Singular Plural Unspecified

She,her, he, him, his, it

We, us, they, them

you


Person and Case Agreement

First Second

Third

Nominative I,we you he,she,they

Accusative me,us you Him,her,them

Genitive my,our

your His,her, their


Gender Agreement

*John has a coffee machine. She loves it.


Syntactic Agreement

Reflexives (himself, herself…) have strong constraints on what syntactic positions they can appear inJohn bought himself a cup of coffee.

*John bought him a cup of coffee.


Selectional Constraints

Jim bought a coffee from the store. He drank it quickly.


Also : Preferences

Recency Grammatical Role Repeated Mention Parallelism Verb Semantics

Based on Salience


Recency

John had a pop-tart. Bill had a jelly donut. Mary wanted it.

Recent Entities are more salient


Grammatical Role

“Sue bought a cup of coffee and a donut from Jane. She met John as she left.”

Entities in subject position are more salient


Repeated Mention

John went to the store to buy coffee. He loves coffee. He drinks 5 cups a day. At the store, Bill sold him a cup. He was delighed.

Entities mentioned more frequently are more salient


Parallelism

John bought coffee from Jim in the morning. Sue bought coffee from him in the evening.

Even with preferences to the contrary (grammatical role) the syntactic parallelism strongly prefers [him = Jim]


Verb Semantics

John telephoned Bill. He was jonesing for coffee.John criticized Bill. He was jonesing for coffee.

Perhaps salience of different elements in the sentence changes with respect to the verb used.


Algorithms --- How to integrate these preferences?

Constraints are easy to use : reject all hypothesis which violate the hard constraints

(if you can accurately detect the constraints!)

Preferences more difficult – how can one integrate these different preferences?


Hobbs Tree Search Algorithm

Given parse trees, search them in a specific order to find the most likely referent


Hobbs in Detail

1. Begin at NP

2. Go up tree to first NP or S. Call this X, and the path p.

3. Traverse all branches below X to the left of p. Propose as antecedent any NP that has a NP or S between it and X

4. If X is the highest S in the sentence, traverse the parse trees of the previous sentences in the order of recency. Traverse left-to-right, breadth first. When a NP is encountered, propose as antecedent. If not the highest node, go to step 5.


Hobbs cont.5. From node X, go up the tree to the first NP or S.

Call it X, and the path p.6. If X is an NP and the path to X did not pass through

the nominal that X dominates, propose X as antecedent

7. Traverse all branches below X to the right of the path, in a left-to-right, breadth first manner. Propose any NP encountered as the antecdent

8. If X is an S node, traverse all brnaches of X to the right of the path but do not go below any NP or S encountered. Propose any NP as the antecedent.


Lappin and Leass (1994) Anaphora Resolution Algorithm

The Lappin and Leass(1994) anaphora resolution algorithm uses

salience weight in determining the antecedent to the pronominals.

It requires as input a fully parsed sentence structure and

uses hierarchy in identifying the subject, object etc.

This algorithm uses syntactic criteria to rule out noun

phrases that cannot possibly corefer with it.

The antecedent is then chosen according to a ranking based

on salience weights.


The salience Factors and WeightsA pronoun P is non-coreferential with a (non-reflexive or non-

reciprocal) noun phrase N if any of the following conditions hold:

P and N have incompatible agreement features. P is in the argument domain of N. P is in the adjunct domain of N. P is an argument of a head

H, N is not a pronoun, and N is contained in H. P is in the NP domain of N. P is a determiner of a noun Q, and N is contained in Q.


Examples

Condition 1:The woman said that he is funny.

Condition 2:She likes her. John seems to want to see him.

Condition 3:She sat near her.

Condition 4:He believes that the man is amusing.This is the man he said John wrote about.

Condition 5:John’s portrait of him is interesting.


Salience Factors and Weights

Salience factor types with initial weightsFactor type Initial weightSentence recency 100Subject emphasis 80Existential emphasis 70Accusative emphasis 50Indirect object and oblique complement emphasis 40Head noun emphasis 80

Non-adverbial emphasis 50


Kennedy 1996The linguistic analysis for anaphora resolution includes

The output of a part of speech tagger,

Augmented with syntactic function annotations for each input token;

Using LINGSOFT


A set of patterns are used for identifying

The NP Chunking with position of the NP in the text: Nominal Sequencing in two subordinate syntactic

environments:a. in an adverbial adjunct b. in an NP (i.e. containment in a prepositional

or clausal complement of a noun, or containment in a relative clause)

Expletive “it”:


Anaphora Resolution

Uses Lappin and Lease algorithmSENT-S: 100 iff in the current sentenceCNTX-S: 50 iff in the current contextSUBJ-S: 80 iff GFUN = subjectEXST-S: 70 iff in an existential constructionPOSS-S: 65 iff GFUN = possessiveACC-S: 50 iff GFUN = direct objectDAT-S: 40 iff GFUN = indirect objectOBLQ-S: 30 iff the complement of a prepositionHEAD-S: 80 iff EMBED = NILARG-S: 50 iff ADJUNCT = NIL


Mitkov 1997

No Parsing of the Input Sentence

Boosting indicators

First Noun Phrases: A score of +1 is assigned to the first NP in a sentence.

Indicating Verbs: A score of +1 is assigned to those NPs immediately following a verb which is a member of a predefined set (including verbs such as discuss, present, illustrate, identify, summarise, examine, describe, define, show, check, develop, review,


MARS Cont….

Lexical Reiteration: A score of +2 is assigned to those NPs repeated twice or more in the paragraph in which the pronoun appears, a score of +1 is assigned to those NPs repeated once in that paragraph.

Section Heading Preference: A score of +1 is assigned to those NPs that also occur in the heading of the section in which the pronoun appears.


Boosting indicators contd..

Collocation Match: A score of +2 is assigned to those NPs that have an identical collocation pattern to the pronoun.

Immediate Reference: A score of +2 is assigned to those NPs appearing in constructions of the form

“… (You) V1 NP … con (you) V2 it (con (you) V3 it)”, where con Є {and/or/before/after…}.

Sequential Instructions: A score of +2 is applied to NPs in the NP1 position of constructions of the form: “To V1 NP1 V2 NP2. (Sentence). To V3 it, V4 NP4“ the noun phrase NP1 is the likely antecedent of the anaphor it (NP1 is assigned a score of 2).

Term Preference: A score of +1 is applied to those NPs identified as representing terms in the genre of the text.


Impeding indicators

Indefiniteness: Indefinite NPs are assigned a score of -1.

Prepositional Noun Phrases: NPs appearing in prepositional phrases are assigned a score of -1.


Indian Language

Types of anaphors Dravidian type with gender marked

pronouns avan, aval, atu (example) Telugu does not

have all other Dravidian languages have. Aryan Language Type without gender

marking in the pronouns “us” as in Hindi (example) FROM CHAPTER


<ANT 1,2,3> uutti </ANT> oru alakiya malai nakaram. Ooty(N) one(Q) beautiful(ADJ) hill(N) town(N). (Ooty is a beautiful hill station.)<ANA 1> ithu </ANA> malaikalin raani. This hill(N)+GEN queen(N). (This is Queen of hills.)<ANA 2> anku </ANA> puungkaa, pataku cavaari, malai rayil untu. There park(N), boating(VBN) hill train(N)

present(V+PRESENT+3S) (There are park, boating, hill train )<ANA 3> athu </ANA> oru cirantha currulaath thalam. it(PN) one(Q) best(ADJ) tourist(N) place(N). (It is a best tourist place.)


“Vasisth” a Rule Based Anaphora Resolution System

1. mo:han(i) avanRe(i) kuttiye kantu. mohan he-poss child-acc see-pst (Mohan saw his child.)2. mo:han(i) avanRe(i) kuttiye kantu ennu kRisnan paRannu. mohan he-poss child-acc see-pst compl krishnan say-pst (Krishnan said that Mohan saw his child.)3. *mo:han(i) avane(i) aticcu. mohan he-acc beat-pst (Mohan beat him.) 4. mo:han avane(i) aticcu ennu kRisnan(i) paRannu. mohan he-acc beat-pst compl krishnan say-pst (Krishnan said that Mohan beat him.)


The Algorithm for Intra-sentential Anaphora

A pronoun P is coreferential with an NP iff the following conditions hold:

a. P and NP have compatible P, N, G features. b. P does not precede NP. c. If P is possessive, then NP is the subject of the clause which contains P. d. If P is non-possessive, then NP is the subject

of the immediate clause which does not contain P.


Vasisth is a multilingual Anaphora Resolution system

Rule based With minimum Parsing Exploit the Morphology of Indian

Languages


“VASISTH” Using Salience Measure for Indian Languages

No In-depth Parsing

Exploit the Rich Morphology of the Language

The analysis depends on the salience weight of the candidate (NP) for the antecedent-hood of an anaphor from a list of probable candidates.


The salience weight assignment

a) The current sentence gets a score of 50 and it reduces by 10 for each preceding sentence till it reaches the fifth sentence. The system considers five sentences for identifying the antecedent.

b) The current clause gets a score of 75 if the pronoun present in the clause is a possessive pronoun and if it is a non-possessive pronoun it gets zero score.

c) The immediate clause gets the score 70 in the case of Possessive pronoun and gets a score of 75 for non-possessive pronouns.

d) For non-immediate clause, the possessive pronoun gets a score of 30 and non-possessive pronoun gets a score of 65.


e)The analysis showed that the subject could be the most probable antecedent for the pronoun. The case markings the subject of a sentence could take are nominative and dative.

A Nominative, a Dative and a Possessive NP with a nominative/Dative head could become a subject of a sentence.


f) The direct object of a sentence could be identified by the case markings and all the case markings other than the subject are considered for object. The next most probable NP for antecedent-hood is the direct object and hence it gets a score of 40.

g) The third NP in a clause, which is not identified as the subject or object, is considered as the indirect object and gets a low score of 30.


Salience factor weights for Indian Languages

Salience Factors Weights

Current sentence Possessive Current clauseImmediate clauseNon-immediate clauseNon-PossessiveCurrent clauseImmediate clauseNon-immediate clausePossessive and Non-PossessiveN.NomN.PossN.DatN.Acc, Loc, Instr…N.others(3rd NP)

50- Reduced by 10 for preceding sentences upto 5th sentence 75 7030 07565 8050504030


How it works

The salience weight to an NP is assigned in the following way

Identify the Pronoun

Consider Four sentences above the sentence containing the Pronoun

Consider all the NPs preceding the Pronoun ( This is the general rule)


Here we take some NPs which follow the the Pronoun since Tamil

All Indian languages are relatively free word Order

Assign Salience Weights.

The NP which gets the maximum salience weight and agrees in png with the anaphor is considered as the antecedent to the anaphor


Machine Learning

CONLL Task and its results


Introduction

Goal of the task Automatically identify coreference chains in a document The coreference chains can include

Names Nominal mentions Pronouns verbs that are coreferenced with a noun phrases.

The data used is the Ontonotes English documents from Ontonotes This consists of five different types of genres

Newswire (NW), Broadcast News (BN), Telephonic Conversation (TC), Broadcast conversation (BC), Web blogs (WB)


Introduction

Coreferencing Two referential entities when they exist in the real

world Coreference analysis determines whether or not two

entities refer to the same entity Coreferents are classified into two types

Pronominal referents Non-pronominal referents

Pronominal referents are pronouns, which refer to other nouns in the text

Non-pronominal referents are names, nominal mentions and other noun phrases


A Sample Text

Eagle Clothes Inc. , which is operating under Chapter 11 of the Federal Bankruptcy Code , said it reached an agreement with its creditors .Under the accord , Albert Roth , chairman and chief executive officer , and Arthur Chase , Sam Beigel , and Louis Polsky will resign as officers and directors of the menswear retailer .Mr. Roth , who has been on leave from his posts , will be succeeded by Geoffrie D. Lurie of GDL Management Inc. , which is Eagle 's crisis manager .Mr. Lurie is currently co-chief executive.


Coreference Tagged Text

<Coref_chain Id="1">Eagle Clothes Inc.</Coref_chainId="1"> , which is operating under Chapter 11 of the federal Bankruptcy Code , said <Coref_chain Id="1">it</Coref_chain Id="1"> reached an agreement with <Coref_chain Id="1">its</Coref_chain Id="1"> creditors .Under the accord , <Coref_chain Id="2">Albert Roth , chairman and chief executive officer</Coref_chain Id=“2"> , and Arthur Chase , Sam Beigel , and Louis Polsky will resign as officers and directors of the <Coref_chain Id="1">menswear retailer</Coref_chain Id="1"> .<Coref_chain Id="2">Mr. Roth , who has been on leave from his posts ,</Coref_chain Id="2"> will be succeeded by <Coref_chain Id="3">Geoffrie D. Lurie of GDL Management Inc. , which is <Coref_chain Id="1">Eagle 's</Coref_chain Id="1"> crisis manager </Coref_chain Id="3">.<Coref_chain Id="3">Mr. Lurie </Coref_chain Id="3"> is currently co-chief executive .


Our Approach Two different approaches are used for the two

types of coreferents Mainly two modules

Pronominal resolution module Non-pronominal resolution module

Pronominal resolution refers to identification of a Noun phrase (NP) that is referred by a pronominal

Non-Pronominal resolution refers to identification of a NP referring to other NP.


Pronominal Resolution Module

Pronominal Resolution is done using refined salience measure.

All pronouns do not refer back to an entity For example in the sentence “It will rain today”

The pronoun “It” does not refer to any entity Such an instance of “It” is called as pleonastic “it”

Before we do pronominal resolution, we need to identify such non-anaphoric pronouns


Pronominal Resolution Module (Contd…)

The first step in the pronominal resolution module is identification of non-anaphoric pronouns and filter the non-anaphoric pronouns

from text The identification of non-anaphoric

pronouns is done using CRFs, a machine learning approach



Using CRFs we build a language model The features used for the training are

Word Part-of-speech (POS) tag

Window of five words Non-anaphoric pronoun filtering has

been observed to improve the results



The second step - core task of the pronoun resolution module the identification of antecedent for a pronoun

Use a refined salience measure based approach

WordNet and NE tag information are used to match the category of pronoun and the antecedent


Pronominal Resolution Module (Contd…) The table below shows the refined salience

factors and the weights assigned for themSalience Factors Weights

Current Sentence (sentence in which pronoun occurs) 100

For the preceding sentences up to four sentences from the current sentence Reduce sentence score by 10

Current Clause (clause in which pronoun occurs) 100 – for possessive pronoun50 – for non-possessive pronouns

Immediate Clause (clause preceding or following the current clause) 50 – for possessive pronoun100 – for non-possessive pronouns

Non-immediate Clause (neither the current or immediate clause) 50

Possessive NP 65

Existential NP 70

Subject 80

Direct Object 50

Indirect Object 40

Compliment of PP 30


Non-Pronominal Resolution Module

Use CRFs, a machine learning method

Preparing the Training data for CRFs learning

All positive pairs of NPs, and negative pairs of NPs are taken for training

Positive pairs are the NPs and the anaphor All NPs between an anaphor and antecedent are the

negative NPs Do not consider the NPs containing the pronouns, We consider the NP on the left side as antecedent NP

and NP on the right side as anaphor NP


Features – Non-pronominal Resolution Module

The features used for training are Distance feature

the possible values are 0,1,2,…. Calculated based on number of sentences between NPs

Definite NP If the antecedent NP is a definite NP, has value 1 else 0

Demonstrative NP If the antecedent NP is demonstrative, has value 1 else 0

String match the possible values are between 0 and 1 calculated as ratio of the number of words matched between

the NPs and the total number of words of the anaphor NP.


Features – Non-pronominal Resolution Module (Contd…)

Number Agreement use the gender data file (Bergsma and Lin, 2006) provided by CoNLL Also use POS information

Gender agreement use the gender data file (Bergsma and Lin, 2006) provided by CoNLL

Alias feature the possible values are 0 or 1. this is obtained using three methods

Comparing the head of the NPs, if both are same then scored as 1 If both the NPs start with NNP or NNPS POS tags, and if they are

same then scored as 1 Looks for Acronym match, if one is an acronym of other it is scored

as 1 Proper NP

If both NPs are proper NPs then value of 1 else 0 NE Tag Info

If NE tag present then value of 1 else 0


Non-Pronominal Resolution Module (Contd…)

The semantic class information (noun category) obtained from the WordNet is used for the filtering purpose. The pairs which do not have semantic

feature match are filtered out.


Complete Coreference Chain Building

The Coreferring pairs obtained from pronominal resolution system and Non-pronominal system are merged to generate the complete coreference chains.

The merging is done as follows: A member of a coreference pair is compared with all the

members of the coreference pairs identified if it occurs in anyone of the pair, then the two pairs are grouped this process is done for all the members of the identified pairs

and the members in each group are aligned based on their position in the document to form the chain

We show the sample outputs for building chain in the next few slides


Evaluation

The data used for training and testing is the English portion of the Ontonotes

The training and development data consisted of 1876 files from different genres viz., Newswire, Broadcast news, Broadcast conversation, Web blogs, magzine articles

The metrics used for evaluating the complete system are MUC B-Cubed CEAFE


Evaluation

We have performed evaluation of the Pronominal resolution module separately using the development data

We perform the evaluation of non-anaphor detection engine In the evaluation of pronominal resolution

module we studied how the non-anaphoric pronoun detection improves


Results – Pronominal Resolution Module

Type of pronoun

Actual (gold standard)

System identified Correctly

Accuracy (%)

Anaphoric Pronouns

939 908 96.6

Non-anaphoric pronouns

160 81 50.6

Total 1099 989 89.9

Evaluation of Non-anaphoric pronoun detection component

System type Total Anaphoric Pronouns

System identified pronouns

System correctly Resolved Pronouns

Prec (%)

Without non-anaphoric pronoun detection

939 1099 693 63.1

With non-anaphoric pronoun detection

939 987 693 70.2

Evaluation of Pronominal resolution module


Results– Complete System

Metric Mention Detection Coreference Resolution

Recall Precision F1 Recall Precision F1

MUC 68.1 61.5 64.6 52.1 49.9 50.9

B-CUBED 68.1 61.5 64.6 66.6 67.6 67.1

CEAFE 68.1 61.5 64.6 42.8 44.9 43.8


Results Discussion

On analysis of the output we found mainly three types of errors. They are Newly invented chains

The system identifies new chains that are not found in the gold standard annotation.

This reduces the precision of the system. This is because of the string match as one of

the features.


Results Discussion (Contd…)

Only head nouns in the chain system while selecting pair for identifying

coreference, the pair has only the head noun instead of the full phrase.

In the phrase “the letters sent in recent days”, the system identifies “the letters” instead of the whole phrase.

This affects both the precision and recall of the system.


Conclusion

Presented a coreference resolution system which combines the pronominal resolution using refined salience based approach with non-pronominal resolution using CRFs, machine learning approach.

Non-anaphoric pronouns identification improves the precision.

In non-pronominal resolution algorithm, the string match feature is an effective feature. But, this feature is found to introduce errors. We need to add additional contextual and semantic feature to

reduce above said errors. The results on the development set are encouraging.


Problems in Anaphora Resolution

Ambiguous Sentences We gave the bananas to the monkeys because

they were hungry. We gave the bananas to the monkeys because

they were ripe. We gave the bananas to the monkeys because

they were here. Complement anaphora

(1) Only a few of the children ate their ice-cream. They ate the strawberry flavour first.

(2) Only a few of the children ate their ice-cream. They threw it around the room instead


The Prime Minister of New Zealand visited us yesterday. The visit was the first time she had come to New York since 1998.

If the second sentence is quoted by itself, it is necessary to resolve the anaphor: The visit was the first time the Prime Minister of New

Zealand had come to New York since 1998. Although of course, as The Prime Minister of New Zealand

is an office of state and she would seem to refer to the person currently occupying that office, it could quite easily be that the Prime Minister of New Zealand had visited New York since 1998 and before the present day, whilst the present incumbent she had not.


Corpus Creation for ML Method

Annotation Guide lines PALINKA MUC GATE Give examples


Tools

GATE Java-RAP (pronouns) GUITAR (Poesio & Kabadjov, 2004;

Kabadjov, 2007) BART (Versleyet al, 2008)


TASKS and Conferences

CoNLL shared task Conference Discourse Anaphora and Anaphora

Resolution Colloquium: DAARC


Where it is required?

Machine Translation Information Extraction Summarization And in……….almost all NLU applications


References

Massimo Poesio Slides: “Anaphora resolution for Practical task”

Ruslan Mitkov: “MARS a Knowledge Poor anaphora resolution system”

MORE FROM OTHER SOURSES


Thank You

Documents

Anaphora Resolution Sobha Lalitha Devi AU-KBC Research Centre MIT Campus of Anna University Chennai-44 [email protected]