1. TEXT PROCESSING 1 Anaphora resolution Introduction to Anaphora Resolution

2. Outline A reminder: anaphora resolution, factors affecting the interpretation of anaphoric expressions A brief history of anaphora resolution  First algorithms: Charniak, Winograd, Wilks  Pronouns: Hobbs  Salience: S-List, LRC Early ML work Definite descriptions: Vieira & Poesio The MUC initiative – also: coreference evaluation methods Soon et al

3. Anaphora resolution: a specification of the problem

4. Anaphora resolution: coreference chains

5. Reminder: Factors that affect the interpretation of anaphoric expressions Factors:  Morphological features (agreement)  Syntactic information  Salience  Lexical and commonsense knowledge Distinction often made between CONSTRAINTS and PREFERENCES

6. Agreement GENDER strong CONSTRAINT for pronouns (in other languages: for other anaphors as well)  [Jane] blamed [Bill] because HE spilt the coffee (Ehrlich, Garnham e.a, Arnold e.a) NUMBER also strong constraint  [[Union] representatives] told [the CEO] that THEY couldn’t be reached

7. Some complexities Gender:  [India] withdrew HER ambassador from the Commonwealth  “…to get a customer’s 1100 parcel-a-week load to its doorstep” [actual error from LRC algorithm] Number:  The Union said that THEY would withdraw from negotations until further notice.

8. Syntactic information BINDING constraints  [John] likes HIM EMBEDDING constraints  [[his] friend] PARALLELISM preferences  Around 60% of pronouns occur in subject position; around 70% of those refer to antecedents in subject position  [John] gave [Bill] a book, and [Fred] gave HIM a pencil Effect of syntax on SALIENCE (Next)

9. Salience In every discourse, certain entities are more PROMINENT

10. Factors that affect prominence Distance Order of mention in the sentence  Entities mentioned earlier in the sentence more prominent Type of NP (proper names > other types of NPs) Number of mentions Syntactic position (subj > other GF, matrix > embedded) Semantic role (‘implicit causality’ theories) Discourse structure

11. Focusing theories Hypothesis: One or more entities in the discourse are the FOCUS OF (LINGUISTIC) ATTENTION just like some entities in the visual space are the focus of VISUAL attention  Grosz 1977, Reichman 1985: ‘focus spaces’  Sidner 1979, Sanford & Garrod 1981: ‘focused entities’  Grosz et al 1981, 1983, 1995: Centering

12. Lexical and commonsense knowledge [The city council] refused [the women] a permit because they feared violence. [The city council] refused [the women] a permit because they advocated violence. Winograd (1974), Sidner (1979) BRISBANE – a terrific right rip from [Hector Thompson] dropped [Ross Eadie] at Sandgate on Friday night and won him the Australian welterweight boxing title. (Hirst, 1981)

13. Problems to be resolved by an AR system: mention identification Effect: recall Typical problems:  Nested NPs (possessives) [a city] 's [computer system]  [[a city]’s computer system]  Appositions: [Madras], [India]  [Madras, [India]]  Attachments

14. Problems for AR: Complex attachments [The quality that’s coming out of [software from [India]]  The quality that’s coming out of software from India is now exceeding the quality of software that’s coming out from the United States scanning through millions of lines of computer code ACE/bnews/devel/ABC19981001.1830.1257

15. Problems for AR: agreement extraction The committee are meeting / is meeting The Union sent a representative. They …. The doctor came to visit my father. SHE told him …

16. Problems to be solved: anaphoricity determination Expletives:  IT’s not easy to find a solution  Is THERE any reason to be optimistic at all? Non-anaphoric definites

17. Outline A reminder: anaphora resolution, factors affecting the interpretation of anaphoric expressions A brief history of anaphora resolution  First algorithms: Charniak, Winograd, Wilks  Pronouns: Hobbs  Salience: S-List, LRC Early ML work Definite descriptions: Vieira & Poesio The MUC initiative, coreference evaluation methods Soon et al 2001

18. A brief history of early work on anaphora resolution Algorithms heavily based on (hand- coded) commonsense knowledge: Charniak 72, Winograd 74, Hobbs 1974, Wilks 1975 Syntax-based algorithms: Hobbs 1974 Salience-based algorithms: Brennan et al, Strube & Hahn, Tetreault

19. First approaches to anaphora resolution: Charniak, Winograd, Wilks Heavy emphasis on inference Hand-coded No formal evaluation

20. 1970s Example: SHRDLU http://hci.stanford.edu/~winograd/shrdlu/ Person: PICK UP A BIG RED BLOCK. Computer: OK. (does it) Person: GRASP THE PYRAMID. Computer: I DON'T UNDERSTAND WHICH PYRAMID YOU MEAN. Person: FIND A BLOCK WHICH IS TALLER THAN THE ONE YOU ARE HOLDING AND PUT IT INTO THE BOX. Computer: BY "IT", I ASSUME YOU MEAN THE BLOCK WHICH IS TALLER THAN THE ONE I AM HOLDING. Computer: OK. (does it) Person: WHAT DOES THE BOX CONTAIN? Computer: THE BLUE PYRAMID AND THE BLUE BLOCK. Person: WHAT IS THE PYRAMID SUPPORTED BY? Computer: THE BOX. Person: HOW MANY BLOCKS ARE NOT IN THE BOX? Computer: FOUR OF THEM. Person: IS AT LEAST ONE OF THEM NARROWER THAN THE ONE WHICH I TOLD YOU TO PICK UP? Computer: YES, THE RED CUBE. This is one of the only dialogs it knows. SHRDLU is too stupid to make mistakes. Beautiful “Demo-ware” Terry Winograd. 1971. MIT Ph.D. Thesis. Terry Winograd

21. Anaphora in SHRDLU First example of HISTORY LIST algorithm Uses a combination of agreement features and semantic constraints Check all possibilities and assign a global score rather than simply find the first match Score incorporates syn component: entities in subj position higher score than entities in object position, in turn ranked more highly than entities in adjunct position Performance made more impressive by including solutions to a number of complex cases, such as reference to events (Why did you do it?) – often ad hoc

22. Hobbs’ `Naïve Algorithm’ (Hobbs, 1974) The reference algorithm for PRONOUN resolution (until Soon et al it was the standard baseline)  Interesting since Hobbs himself in the 1974 paper suggests that this algorithm is very limited (and proposes one based on semantics) The first anaphora resolution algorithm to have an (informal) evaluation Purely syntax based

23. Hobbs’ `Naïve Algorithm’ (Hobbs, 1974) Works off ‘surface parse tree’ Starting from the position of the pronoun in the surface tree,  first go up the tree looking for an antecedent in the current sentence (left- to-right, breadth-first);  then go to the previous sentence, again traversing left-to-right, breadth-first.  And keep going back

24. Hobbs’ algorithm: Intrasentential anaphora Steps 2 and 3 deal with intrasentential anaphora and incorporate basic syntactic constraints: Also: John’s portrait of him S NP John V likes NP him X p

25. Hobbs’ Algorithm: intersentential anaphora S NP John V likes NP him S NP Bill V is NP a good friend X candidate

26. Evaluation The first anaphora resolution algorithm to be evaluated in a systematic manner, and still often used as baseline (hard to beat!) Hobbs, 1974:  300 pronouns from texts in three different styles (a fiction book, a non- fiction book, a magazine)  Results: 88.3% correct without selectional constraints, 91.7% with SR  132 ambiguous pronouns; 98 correctly resolved. Tetreault 2001 (no selectional restrictions; all pronouns)  1298 out of 1500 pronouns from 195 NYT articles (76.8% correct)  74.2% correct intra, 82% inter Main limitations  Reference to propositions excluded  Plurals  Reference to events

27. Salience-based algorithms Common hypotheses:  Entities in discourse model are RANKED by salience  Salience gets continuously updated  Most highly ranked entities are preferred antecedents Variants:  DISCRETE theories (Sidner, Brennan et al, Strube & Hahn): 1-2 entities singled out  CONTINUOUS theories (Alshawi, Lappin & Leass, Strube 1998, LRC): only ranking

28. Salience-based algorithms Sidner 1979:  Most extensive theory of the influence of salience on several types of anaphors  Two FOCI: discourse focus, agent focus  never properly evaluated Brennan et al 1987 (see Walker 1989)  Ranking based on grammatical function  One focus (CB) Strube & Hahn 1999  Ranking based on information status (NP type) S-List (Strube 1998): drop CB  LRC (Tetreault): incremental

29. LRC An update on Strube’s S-LIST algorithm (= Centering without centers) Initial version augmented with various syntactic and discourse constraints

30. LRC Algorithm (LRC) Maintain a stack of entities ranked by grammatical function and sentence order (Subj > DO > IO) Each sentence is represented by a Cf-list: list of salient entities ordered by grammatical function While processing utterance’s entities (left to right) do:  Push entity onto temporary list (Cf-list-new), if pronoun, attempt to resolve first: Search through Cf-list-new (l-to-r) taking the first candidate that meets gender, agreement constraints, etc. If none found, search past utterance’s Cf-lists starting from previous utterance to beginning of discourse

31. Results AlgorithmPTB-News (1694) PTB-Fic (511) LRC74.9%72.1% S-List71.7%66.1% BFP59.4%46.4%

32. Comparison with ML techniques of the time AlgorithmAll 3 rd LRC76.7% Ge et al. (1998)87.5% (*) Morton (2000)79.1%

33. Carter’s Algorithm (1985) The most systematic attempt to produce a system using both salience (Sidner’s theory) & commonsense knowledge (Wilks’s preference semantics) Small-scale evaluation (around 100 hand- constructed examples) Many ideas found their way in the SRI’s Core Language Engine (Alshawi, 1992)

34. Outline A reminder: anaphora resolution, factors affecting the interpretation of anaphoric expressions A brief history of anaphora resolution  First algorithms: Charniak, Winograd, Wilks  Pronouns: Hobbs  Salience: S-List, LRC Modern work in anaphora resolution Early ML work The MUC initiative – also: coreference evaluation methods Soon et al

35. MODERN WORK IN ANAPHORA RESOLUTION Availability of the first anaphorically annotated corpora circa 1993 (MUC6) made statistical methods possible

36. STATISTICAL APPROACHES TO ANAPHORA RESOLUTION UNSUPERVISED approaches  Eg Cardie & Wagstaff 1999, Ng 2008 SUPERVISED approaches  Early (NP type specific)  Soon et al: general classifier + modern architecture

37. ANAPHORA RESOLUTION AS A CLASSIFICATION PROBLEM 1. Classify NP1 and NP2 as coreferential or not 2. Build a complete coreferential chain

38. SOME KEY DECISIONS ENCODING  I.e., what positive and negative instances to generate from the annotated corpus  Eg treat all elements of the coref chain as positive instances, everything else as negative: DECODING  How to use the classifier to choose an antecedent  Some options: ‘sequential’ (stop at the first positive), ‘parallel’ (compare several options)

39. Outline A reminder: anaphora resolution, factors affecting the interpretation of anaphoric expressions A brief history of anaphora resolution  First algorithms: Charniak, Winograd, Wilks  Pronouns: Hobbs  Salience: S-List, LRC Modern work in anaphora resolution Early ML work The MUC initiative – also: coreference evaluation methods Soon et al

40. Early machine-learning approaches Main distinguishing feature: concentrate on a single NP type Both hand-coded and ML:  Aone & Bennett (pronouns)  Vieira & Poesio (definite descriptions) Ge and Charniak (pronouns)

41. Definite descriptions: Vieira & Poesio A first attempt at going beyond pronouns while still doing a large-scale evaluation Definite descriptions chosen because they require lexical and commonsense knowledge Developing both a hand-coded and a ML decision tree (as in Aone and Bennett) Vieira & Poesio 1996, Vieira 1998, Vieira & Poesio 2000

42. Preliminary corpus study (Poesio and Vieira, 1998) Annotators asked to classify about 1,000 definite descriptions from the ACL/DCI corpus (Wall Street Journal texts) into three classes: DIRECT ANAPHORA: a house … the house DISCOURSE-NEW: the belief that ginseng tastes like spinach is more widespread than one would expect BRIDGING DESCRIPTIONS: the flat … the living room; the car … the vehicle

43. Poesio and Vieira, 1998 Results: More than half of the def descriptions are first-mention Subjects didn’t always agree on the classification of an antecedent (bridging descriptions: ~8%)

44. The Vieira / Poesio system for robust definite description resolution Follows a SHALLOW PROCESSING approach (Carter, 1987; Mitkov, 1998): it only uses Structural information (extracted from Penn Treebank) Existing lexical sources (WordNet) (Very little) hand-coded information

45. Methods for resolving direct anaphors DIRECT ANAPHORA: the red car, the car, the blue car: premodification heuristics segmentation: approximated with ‘loose’ windows

46. Methods for resolving discourse- new definite descriptions DISCOURSE-NEW DEFINITES the first man on the Moon, the fact that Ginseng tastes of spinach: a list of the most common functional predicates (fact, result, belief) and modifiers (first, last, only… ) heuristics based on structural information (e.g., establishing relative clauses)

47. The (hand-coded) decision tree 1. Apply ‘safe’ discourse-new recognition heuristics 2. Attempt to resolve as same-head anaphora 3. Attempt to classify as discourse new 4. Attempt to resolve as bridging description. Search backward 1 sentence at a time and apply heuristics in the following order: 1. Named entity recognition heuristics – R=.66, P=.95 2. Heuristics for identifying compound nouns acting as anchors – R=.36 3. Access WordNet – R, P about.28

48. The decision tree obtained via ML Same features as for the hand-coded decision tree Using ID3 classifier (non probabilistic decision tree) Training instances:  Positive: closest annotated antecedent  Negative: all mentions in previous four sentences Decoding: consider all possible antecedents in previous four sentences

49. Automatically learned decision tree

50. Overall Results Evaluated on a test corpus of 464 definite descriptions Overall results: RPF Version 153%76%62% Version 257%70%62% D-N def77% ID375%

51. Overall Results Results for each type of definite description: RPF Direct anaphora 62%83%71% Disc new69%72%70% Bridging29%38%32.9%

52. Outline A reminder: anaphora resolution, factors affecting the interpretation of anaphoric expressions A brief history of anaphora resolution  First algorithms: Charniak, Winograd, Wilks  Pronouns: Hobbs  Salience: S-List, LRC Early ML work Definite descriptions: Vieira & Poesio The MUC initiative and coreference evaluation methods Soon et al

53. MUC First big initiative in Information Extraction Produced first sizeable annotated data for coreference Developed first methods for evaluating systems

54. MUC terminology: MENTION: any markable COREFERENCE CHAIN: a set of mentions referring to an entity KEY: the (annotated) solution (a partition of the mentions into coreference chains) RESPONSE: the coreference chains produced by a system

55. Evaluation of coreference resolution systems Lots of different measures proposed ACCURACY:  Consider a mention correctly resolved if Correctly classified as anaphoric or not anaphoric ‘Right’ antecedent picked up Measures developed for the competitions:  Automatic way of doing the evaluation More realistic measures (Byron, Mitkov)  Accuracy on ‘hard’ cases (e.g., ambiguous pronouns)

56. Vilain et al 1995 The official MUC scorer Based on precision and recall of links

57. Vilain et al: the goal The problem: given that A,B,C and D are part of a coreference chain in the KEY, treat as equivalent the two responses: And as superior to:

58. Vilain et al: RECALL To measure RECALL, look at how each coreference chain S i in the KEY is partitioned in the RESPONSE, and count how many links would be required to recreate the original, then average across all coreference chains.

59. Vilain et al: Example recall In the example above, we have one coreference chain of size 4 (|S| = 4) The incorrect response partitions it in two sets (|p(S)| = 2) R = 4-2 / 4-1 = 2/3

60. Vilain et al: precision Count links that would have to be (incorrectly) added to the key to produce the response I.e., ‘switch around’ key and response in the equation before

61. Beyond Vilain et al Problems:  Only gain points for links. No points gained for correctly recognizing that a particular mention is not anaphoric  All errors are equal Proposals:  Bagga & Baldwin’s B-CUBED algorithm  Luo recent proposal

62. Outline A reminder: anaphora resolution, factors affecting the interpretation of anaphoric expressions A brief history of anaphora resolution  First algorithms: Charniak, Winograd, Wilks  Pronouns: Hobbs  Salience: S-List, LRC Early ML work Definite descriptions: Vieira & Poesio The MUC initiative – also: coreference evaluation methods Soon et al

63. Soon et al 2001 First ‘modern’ ML approach to anaphora resolution  Resolves ALL anaphors  Fully automatic mention identification Developed instance generation & decoding methods used in a lot of work since

64. Preprocessing Cfr. Soon et al

65. Soon et al: preprocessing  POS tagger: HMM-based 96% accuracy  Noun phrase identification module HMM-based Can identify correctly around 85% of mentions (?? 90% ??)  NER: reimplementation of Bikel Schwartz and Weischedel 1999 HMM based 88.9% accuracy

66. Soon et al: training instances

67. Soon et al 2001: Features NP type Distance Agreement Semantic class

68. Soon et al: NP type and distance NP type of antecedent i i-pronoun (bool) NP type of anaphor j (3) j-pronoun, def-np, dem-np (bool) DIST 0, 1, …. Types of both both-proper-name (bool)

69. Soon et al features: string match, agreement, syntactic position STR_MATCH ALIAS dates (1/8 – January 8) person (Bent Simpson / Mr. Simpson) organizations: acronym match (Hewlett Packard / HP) AGREEMENT FEATURES number agreement gender agreement SYNTACTIC PROPERTIES OF ANAPHOR occurs in appositive contruction

70. Soon et al: semantic class agreement PERSON FEMALEMALE OBJECT DATEORGANIZATION TIMEMONEYPERCENT LOCATION SEMCLASS = true iff semclass(i) <= semclass(j) or viceversa

71. Soon et al: generating training instances Marked antecedent used to create positive instance All mentions between anaphor and marked antecedent used to create negative instances

72. Generating training instances ((Eastern Airlines) executives) notified ((union) leaders) that (the carrier) wishes to discuss (selective (wage) reductions) on (Feb 3) POSITIVE NEGATIVE

73. Soon et al: decoding Right to left, consider each antecedent until classifier returns true

74. Decoding example

75. Soon et al: evaluation MUC-6:  P=67.3, R=58.6, F=62.6 MUC-7:  P=65.5, R=56.1, F=60.4

76. Soon et al: evaluation

77. Subsequent developments Different models of the task Different preprocessing techniques Using lexical / commonsense knowledge (particularly semantic role labelling)  Next lecture Salience Anaphoricity detection Development of AR toolkits (GATE, LingPipe, GUITAR)

78. Other models Cardie & Wagstaff: coreference as (unsupervised) clustering  Much lower performance Ng and Cardie Yang ‘twin-candidate’ model

79. Ng and Cardie 2002:  Changes to the model: Positive: first NON PRONOMINAL Decoding: choose MOST HIGH PROBABILITY  Many more features: Many more string features Linguistic features (binding, etc) Subsequently:  Discourse new detection (see below)

80. Readings Kehler’s chapter in Jurafsky & Martin  Alternatively: Elango’s survey http://pages.cs.wisc.edu/~apirak/cs/cs838/pradheep- survey.pdf http://pages.cs.wisc.edu/~apirak/cs/cs838/pradheep- survey.pdf Hobbs J.R. 1978, “Resolving Pronoun References,” Lingua, Vol. 44, pp. 311-. 338.  Also in Readings in Natural Language Processing, Renata Vieira, Massimo Poesio, 2000. An Empirically-based System for Processing Definite Descriptions. Computational Linguistics 26(4): 539- 593 W. M. Soon, H. T. Ng, and D. C. Y. Lim, 2001. A machine learning approach to coreference resolution of noun phrases. Computational Linguistics, 27(4):521--544,