A game-theoretic model of referential coherence and its empirical verification using large Japanese and English corpora

Referential coherence represents the smoothness of discourse resulting from topic continuity and pronominalization. Rational individuals prefer a referentially coherent structure of discourse when they select a language expression and its interpretation. This is a preference for cooperation in communication. By what principle do they share coherent expressions and interpretations? Centering theory is the standard theory of referential coherence [Grosz et al. 1995]. Although it is well designed on the bases of first-order inference rules [Joshi and Kuhn 1979], it does not embody a behavioral principle for the cooperation evident in communication. Hasida [1996] proposed a game-theoretic hypothesis in relation to this issue. We aim to empirically verify Hasida's hypothesis by using corpora of multiple languages. We statistically design language-dependent parameters by using a corpus of the target language. This statistical design enables us to objectively absorb language-specific differences and to verify the universality of Hasida's hypothesis by using corpora. We empirically verified our model by using large Japanese and English corpora. The result proves the language universality of the hypothesis.