Parameter Setting In Centering Theory And Its Effects On Chinese Anaphora Resolution: An Empirical Study

Centering Theory is proved to be an extremely useful theory of discourse coherence and salience. However, the basic concepts in Centering Theory have been left unspecified purposely for its more flexible accommodation of different concrete tasks.Based on the parameter setting research on Centering and various Centering-based anaphora resolutions, this study explores the parameter setting of Centering specifically for Chinese anaphora resolution.For this purpose, we have collected a corpus of 31,493 words from three different discourse genres, among which there are 5148 noun phrases in total. We annotated the noun phrases in the corpus with grammatical functions and morphological features. Then a database of nominal information is generated from the annotated corpus. Six Centering-based resolution algorithms, each realizing an instantiation of parameter setting, are developed in this study to automatically resolve the 1442 zero anaphors and 278 pronouns in the corpus.The Centering parameters examined in this study are identification of utterance, ranking, and Rl-pronouns. The two instantiations for utterance definition adopted in this research are Udef.1, identifying utterances with the discourse sequences containing at least one predicate structure and demarked by punctuations including comma types and full stop types, and Udef.2, identifying utterances with sentences. Our research found that Udef.l, compared with Udef.2, is an ideal utterance definition in zero anaphora resolution. Pronoun resolution comparatively has lower degree of sensitiveness to the change of utterance definitions. The findings of our utterance definition research suggest that zero anaphors, as high accessibility markers, require quicker updating of Centering or focus state to catch the instantly changing discourse attention.The ranking-affecting factors examined in this research are surface order, grammatical functions, parallelism in grammatical roles, C_b continuity and structural hierarchical consideration. The data results of anaphora resolution show that grammatical functions are more delicate and reliable salience indicators for discourse entities than surface order of occurrence of discourse entities. Considering parallelism in grammatical roles is proved empirically in this research to be contributory and useful in the resolution of both zero anaphors and pronouns, and its positive effect would be more significant if it is applied under Udef.2. C_b continuity is proved to be futile in the resolution of both zero anaphors and pronouns. Under certain circumstances, including C_b continuity into ranking even leads to negative effects on the resolution of anaphors. The futility of C_b continuity in...