1. An Entity-Mention Model for Coreference Resolution with Inductive Logic Programming Xiaofeng Yang 1 Jian Su 1 Jun Lang 2 Chew Lim Tan 3 Ting Liu 2 Sheng Li 2 Reporter: Chia-Ying Lee Advisor: Prof. Hsin-Hsi Chen 1 Institute for Infocomm Research 2 Harbin Institute of Technology 3 National University of Singapore ACL 2008

2. 22 Introduction Coreference resolution : The process of linking multiple mentions that refer to the same entity coreference; anaphor 同指涉 antecedent 先行詞 Inductive logic programming : Supervised learning Inductive rule from positive cases 2 2

3. 33 Related Work 1. Mention pair model Aone and Bennett (1995); McCarthy and Lehnert (1995); Soon et al. (2001); Ng and Cardie (2002)) Individual mention usually lacks adequate descriptive information of the referred entity (ex: Powell vs he) 2. Entity-mention model Luo et al., 2004; Yang et al., 2004 Can’t describing each individual mention in an entity 3. Inductive logic programming in NLP Parsing (Mooney, 1997) POS disambiguation (Cussens, 1996) Lexicon construction (Claveau et al., 2003) WSD (Specia et al., 2007) 3

4. 4 Modeling Coreference Resolution The probability that a mention belongs to an entity Example e1 : Microsoft Corp. - its - The company e2 : its new CEO - he e3 : yesterday 4

5. 5 Mention-Pair Model Soon et al. (2001) and Ng and Cardie (2002) Instance i{m k, m j } m j is an active mention & m k is a preceding mention Positive: m j and its closest antecedent (only one for m j ) Negative: every intervening mentions between mj and its closest antecedent m j is linked with the mention that is classified as positive (if any) with the highest confidence value

6. 6 Feature Set for Coreference Resolution 123123 同位語 述詞

7. 7 Entity-Mention Model Mention-pair model error: Lack adequate descriptive information “ Mr. Powell ”, “ Powell ”, and “s he ” Instance i{m k, m j } Positive: m j and the entity to which m j belongs. Negative: every entity whose last mention occurs between m j and its closest antecedent If no positive entity exists, the m j forms a new entity entity features: first-order features Any-X, Most-X, All-X Distance feature :the minimum distance between the mentions in the entity and the active mention.

8. 8 Entity-mention Model with ILP (1/3) Tool: ALEPH by Srinivasan (2000) (Oxford) Input: positive example E + negative example E - background knowledge K Output: hypotheses h e 1_6 denotes the part of e 1 before m 6, example representation: link(e 1_6, m 6 ) 8

9. 9 Entity-mention Model with ILP (2/3) background knowledge K predicates 1. Information related to e i_j and m j 2. Relations between e i_j and its mentions has_mention(e 1_6, m 6 ) 3. Information related to m j and each mention m k in e i_j 9

10. 10 Entity-mention Model with ILP (3/3) Hypothesis rule link(A,B) :- has mention(A,C), numAgree(B,C,1), strMatch Head(B,C,1), bareNP(C,1).

11. 11 Experiments and Result(1/4) Corpus: ACE-2 V1.0 corpus (NIST, 2003) Modify ILP tool, ALEPH: Rule accuracy 100% to 50% 3 predicates to 10 predicates 11

12. 12 Baseline model: C4.5 algorithm Preprocessing Tokenizer Part-of-Speech tagger accuracy of 97% on Penn WSJ TreeBank NP chunker (Zhou and Su, 2000) F-measure above 94% on Penn WSJ TreeBank Named-Entity Recognizer (Zhou and Su, 2002) F-measure of 96.6% (MUC-6) and 94.1%(MUC-7) 12 Experiments and Result(2/4)

13. 13 Experiments and Result(3/4) F-measure is 2-4% lower than the state-of-the- art, which utilized sophisticated semantic or real world knowledge Significant under 2-tailed t test (p < 0.05) 13

14. 14 Experiments and Result(4/4) Multiple non-instantiated arguments (i.e. C and D) could possibly appear in the same rule

15. 15 Conclusion & Future Work The model can express the relations between an entity and its mentions, and to automatically learn the first-order rules ILP based entity-model performs better than the mention-pair model (with up to 2.3% increase in F- measure) Future work: Investigate more sophisticated clustering methods that would lead to global optimization keeping a large search space (Luo et al., 2004) using integer programming (Denis and Baldridge, 2007) 15

16. 16 Thank you! 16