Clause Anaphora Resolution for Japanese Demonstrative Determiner based on Semantic Similarity between Different Parts-of-Speech

Clause Anaphora Resolution for Japanese Demonstrative Determiner based on Semantic Similarity between Different Parts-of-Speech Yukino Ikegami Ernesto Damiani Akihiro Urano Setsuo Tsuruta At SMC 2013

Background • Informal/colloquial style texts often contain anaphoric expressions (e.g.) I like such words. • What is “such” in “such words” means? ➢ The answer depends on context ➢ There is an ambiguity ➢ Anaphora resolution: Automatically resolving anaphoric relation 2013/10/15 SMC 2013 2

Related works • Conventional anaphora resolutions mainly deal with nominal anaphoric • Heuristic rule & ontology based anaphora resolution [Murata 2000] – Resolves {nominal, adjective, adverb} anaphoric – It assumes that antecedent is noun phrase or previous sentence 2013/10/15 SMC 2013 3

Problems on Anaphora Resolution • Conventional anaphora resolutions are not enough to resolve following case: ◼Demonstrative determiner refers a clause – We call “clause anaphoric” – E.g. He inquire xxx. I think such question ---. – Part-of-speech (POS) of anaphor and POS of antecedent are often different in this case 2013/10/15 SMC 2013 4

Our proposed method • Semantic similarity based anaphora resolution for Japanese demonstrative determiner • Consider synonymous relationship crossing parts-of-speech – E.g. Deal with the verb “talk” as synonym of the noun “tale” 2013/10/15 SMC 2013 5

Procedure of our method • Preprocessing phase 1. Dependency Parsing 2. Semantic Role Labeling & Giving word semantics 3. Conceptual Dependency structure parsing • Anaphora Resolution phase 1. Find demonstrative pronouns 2. Extend semantics of anaphor words using Semantically relationship table 3. Choose the most similar candidate 2013/10/15 SMC 2013 6

Preprocessing 1. Syntactic dependency structure parsing – CaboCha [Kudo et al. 2002] 2. Semantic role labeling & giving word semantics – ASA [Takeuchi et al. 2010] 3. Conceptual Dependency Structure Parsing – Conceptual dependency [Schank 1972] 2013/10/15 SMC 2013 7

Anaphora resolution (1) Find Demonstrative Determiner • Finds demonstrative determiner by morphological analysis (e.g.) I wonder what that story told by the managing director was. • If not found, terminates anaphora resolution 2013/10/15 SMC 2013 8

Anaphora resolution (2) Extend word semantics 1. Extract similar words from semantically relationship table 言う (say) - 発声 (utterance) Semantic relation Table 言う (say) – 話 (story) 話 (story) - 発言(statement) - 言葉 (word) - 発声 (utterance) Add 1. Add semantics of similar words to semantics of the anaphor phrase 2013/10/15 SMC 2013 9

Anaphora resolution (3) Choose the most similar candidate • Measure semantic similarity between anaphor phrase and each candidates • Ontology path similarity Lm Ln + L1 L2 sim ( w1, w2 ) = Lm + Ln • The most similar word is chosen as antecedent 2013/10/15 SMC 2013 10

Evaluation Experiment Counseling dialogue Accuracy 100% Twitter 63.04% Novel 62.5% Dataset: Contain the phrase あの話 (“that conversation”) Accuracy: Agreement rate between human and our method 2013/10/15 SMC 2013 11

Major causes of failures in evaluation experiments • Lack of synonymous relationship – Need automatically acquiring synonymous relationship • Specific expression in written language – E.g. Use quotation marks instead of verbs – Need cooperating with rule based method • Cataphoric/exophoric expression – Need detecting implied reference 2013/10/15 SMC 2013 12

Conclusion • Semantic similarity based anaphora resolution for Japanese demonstrative determiners – Consider synonymous relationship crossing Parts-of-speech (POS) – Can resolve clause anaphoric, not only nominal anaphoric 2013/10/15 SMC 2013 13

References • [Murata 2000] M. Murata, “ Anaphora resolution in japanese sentences using surface expressions and examples, ” arXiv:preprint, cs/0009011, 2000. • [Kudo et al. 2002] T.Kudo and Y.Matsumoto. “Japanese dependency analysis using cascaded chunking”, in Proc. of the 6th Conference on Natural Language Learning 2002, pp. 63-69, 2002. • [Takeuchi et al. 2010] K. Takeuchi, S. Tsuchiyama, M. Moriya and Y. Moriyasu, “ Construction of argument structure analyzer toward searching same situations and actions ” (in Japanese), IEICE technical report Natural language understanding and models of communication, vol. 109, No.390, pp.1-6, 2010. • [Schank, 1972] R. Schank. “ Conceptual dependency: a theory of natural language understanding, ” Cognitive Psychology, Vol.3, No.4, 1972. 2013/10/15 SMC 2013 14