Resolving Abstract Definite Anaphora in Chinese Texts

Computational Linguistics and Chinese Language Processing Vol. 16, No. 1-2, March/June 2011, pp. 15-26 15 © The Association for Computational Linguistics and Chinese Language Processing Resolving Abstract Definite Anaphora in Chinese Texts Tyne Liang∗ and Jyun-Hua Cheng∗ Abstract Anaphora is a rhetorical device commonly used in written texts. It denotes the use of terms referring to previously-mentioned entities, concepts, or events. In this paper, the definite anaphora in Chinese texts is addressed and empirical approaches to tackle abstract anaphors are presented. The resolution is built on the association between target anaphors and the corresponding referents in their multiple-type features extracted from different levels of discourse units. Experimental results show that features extracted from clauses are more useful than those extracted from sentences in referent identification. Besides, the presented salience-based model outperforms the SVM-based model no matter whether the best set of extracted features is employed or not. Keywords: Anaphora Resolution, Chinese Text, Definite Anaphora, Feature Extraction 1. Introduction 1.1 Motivation Anaphora is an instance of an expression referring to the preceding utterances. Effective anaphora resolution enhances understanding of a text and facilitates many applications of natural language processing. The resolution involves anaphor recognition and referent recognition. In Chinese texts, anaphors can be missing or be present as pronouns, demonstratives and definite descriptions. Common pronouns are like “他” (“he, him”), “她” (“she, her”), “它” (“it”), “我們” (“we, us”), “他們” (“they, them”); demonstratives are “這” (“this”), ”那” (“that”) and definite description are like the pattern “這+[quantifier ]+noun phrase.” Without concerning zero anaphora, about 54% of the explicit anaphors are pronouns, 40% are definite descriptions, and 6% are demonstratives in a corpus containing 20 news articles. Essentially, the challenges involved with Chinese anaphora resolution are attributed to the complexities of Chinese sentence structures. It is known that although a Chinese sentence ∗ College of Computer Science, National Chiao Tung University, Hsinchu, Taiwan, R.O.C. E-mail: tliang@cs.nctu.edu.tw; sunrise0406.iit97g@g2.nctu.edu.tw 16 Tyne Liang and Jyun-Hua Cheng features the subject-verb-object order, the sentence may be formed by a series of verbs or by pronoun or subject dropping, thus making sentence parsing difficult. Moreover, there is no blank space between adjacent words in Chinese sentences, making word or noun phrase identification difficult. Unlike most previous research projects focusing on non-abstract anaphora resolution, this paper addresses the definite anaphora in Chinese written texts and presents empirical parser-free approaches to resolve abstract anaphors, like ”這項方案” (“this plan”). The resolution is based on the linking between anaphors and their referents in multiple aspects of contextual, semantic and surface features. Among them, semantic features are extracted with the help of three outer resources, namely, Tongyici Cilin1 (TYCC for short), CKIP Lexicon2, and Google search results3. Additionally, the features extracted from different discourse units are investigated and the best feature set is verified at referent identification. In the experiments, both SVM-based and salience models are implemented for model comparison. Experimental results show that the features extracted from clauses are more useful than those extracted from sentences for anaphora resolution. Besides, the presented salience-based model …

Content truncated to 500 words