An Evaluation Of Centering Parameters In Anaphora Resolution

Anaphora study has been a heatedly discussed topic in the linguistic literature, and researchers have carried out extensive studies by using various different linguistic theories. Anaphora resolution, originating from natural language processing and primarily aiming at resolving anaphors in discourse by computer, represents a new direction in anaphora study because its computational linguistic method can contribute more directly to automatic information processing, particularly anaphora production and anaphora interpretation. It has attracted some scholars'attention, but contrastive study in the field is largely neglected.Among the various approaches to anaphora resolution, Centering Theory is one of the most attractive, since it incorporates linguistic theorizing with the practicability of computational operation. Centering Theory is concerned with the relationships among foci of attention, choice of referring expressions and perceived coherence of utterances within a discourse segment. It defines local discourse coherence and salience of discourse entities with the concept of"centers"in terms of three constraints and two rules. Constraint 3 and Rule 1 in particular offer an ideal basis for anaphora resolution. The theory was proposed in early 1980s, and has been widely used in anaphora resolution and text processing in various languages. Our research is based on Centering Theory not only because of its proven effectiveness in anaphora resolution but also because the model it offers is convenient for a contrastive study as we can manipulate the parameters of"utterance","R1-pronoun"and"Cf-ranking"to examine and compare their effects in Chinese and English anaphora resolution.The thesis carries out an empirical study by running six anaphora resolution algorithms on a manually annotated corpus containing about 30,000 words. It tries to find out an optimal configuration of the above three Centering parameters in English and Chinese, examines the effects of each parameter and ranks them by their influence in anaphora resolution, and compares the results in the two languages.Our results show that the influences of the three parameters are different in Chinese and English. In Chinese, the most influential parameter is R1-pronoun, and zero pronouns prove to be typical R1-pronouns whose behavior is in keeping with the predictions of Centering Theory. Next to R1-pronoun is utterance, and to identify utterance with predicate structure gets a good result. The detection of zero topics also significangly improves the results. Ranking, though is the least influential, still plays a crucial role in anaphora resolution, and the detection of zero topics significantly improves the results.In English, R1-pronoun also ranks the first, but third person pronouns, the typical and dominant R1-pronouns in English, don't conform to Centring Theory as well as zero pronouns do. Pronouns in different grammatical positions demonstrate certain divergence in influencing the results of anaphora resolution: those in subject positions conform more to Centering Thoery than those in other positions. Utterance ranks the second, and to identity utterance with tensed clause shows a good result. Besides, main/subordinate distinction plays a significant role in improving the results. The least influential one is Cf-ranking.