Automatic detection of antecedents of Japanese zero pronouns using a Japanese-English bilingual corpus

Automatic Detection of Antecedents of Japanese Zero Pronouns Using a Japanese-English Bilingual Corpus Dong Zhan 1301211064@pku.edu.cn School of Software and Microelectronics, Peking University, Beijing, 100871, China Hiromi Nakaiwa nakaiwa@is.nagoya-u.ac.jp Graduate School of Information Science, Nagoya University, Aichi, 464-8601, Japan Abstract In this paper we present a method of detecting zero pronouns in Japanese clauses and identifying their antecedents using aligned sentence pairs from a Japanese-English bilingual corpus and open resource tools. We use syntactic and semantic structures and the alignment of words and phrases in the sentence pairs to automatically detect zero pronouns and determine their antecedents using English translations. We build rules to link antecedents with zero pronouns and create filters to remove problematic sentence pairs. Experimental results confirm the effectiveness of our method. The proposed method allows the construction of an annotated corpus of zero pronoun sentences in which the antecedents of the missing pronouns are flagged. This would be very useful for machine translation (MT), because zero pronoun detection is a vital problem when translating languages which allow zero pronouns. 1 Introduction In many languages, such as Japanese and Chinese, elements which can be easily de- duced by the reader are frequently omitted from subsequent expressions in discourses. Omissions related to obligatory subject and object cases are often referred to as zero pronouns. A zero pronoun can be thought of as a noun phrase which is obligatory but which is not expressed explicitly. Zero pronoun resolution is more complicated than overt pronoun resolution because a natural language processing (NLP) system has to detect zero pronouns before it can identify their antecedents. Determination of the antecedents of zero pronouns is vital in many NLP applications, including information extraction, question answering, machine translation, etc. Many algorithms have been proposed for pronominal anaphora resolution [9, 2, 10]. However, resolving the anaphora problem is much harder in languages which allow zero pronouns compared to languages which do not allow zero pronouns, such as English, since detection of zero pronouns should be done before anaphora resolution. Several methods have been proposed to solve this problem [12, 3, 14]. Some of these methods use semantic and pragmatic constraints, such as semantic constraints on cases, modal expressions, or verbal semantic attributes, to determine the referents of zero pronouns [12]. Other methods use a machine learning approach for anaphora resolution, such as using semantic attributes as computable features to perform identification and resolution Proceedings of MT Summit XV, vol.1: MT Researchers' Track Miami, Oct 30 - Nov 3, 2015 | p. 66 of anaphoric zero pronouns in Chinese [14] or using syntactic patterns of zero pronouns and their antecedents as features for zero-anaphora resolution in Japanese [3, 5]. Most of these methods focus on zero pronoun resolution and detection of an- tecedents of zero pronouns as a crucial part of zero pronoun resolution. All of these methods employ either an annotated corpus containing zero pronoun sentences, with anaphoric relationships tagged by annotators [3, 4], or an annotated corpus provided by other persons or institutes [14, …

Content truncated to 500 words