Building a Diverse Document Leads Corpus Annotated with Semantic Relations

Copyright 2012 by Masatsugu Hangyo, Daisuke Kawahara, and Sadao Kurohashi 26th Pacific Asia Conference on Language,Information and Computation pages 535–544 Building a Diverse Document Leads Corpus Annotated with Semantic Relations Masatsugu Hangyo Daisuke Kawahara Sadao Kurohashi Graduate School of Informatics, Kyoto University Yoshida-honmachi, Sakyo-ku Kyoto, 606-8501, Japan {hangyo,dk,kuro}@nlp.ist.i.kyoto-u.ac.jp Abstract In these days, semantic analysis has been ac- tively studied in natural language processing. For the study of semantic analysis, corpora with semantic annotations are essential. Al- though there are such corpora annotated on newspaper articles, there are various genres and styles, including linguistic expressions that are not found in newspaper articles. In this paper, we build a diverse document leads corpus annotated with semantic relations. To reduce the workload of annotators and anno- tate as many various documents as possible, we restrict the annotation target of each docu- ment to only the first three sentences. We have completed building a corpus of 1,000 docu- ments and report the statistics of this corpus. 1 Introduction In recent years, semantic analysis including predicate-argument structure analysis and anaphora resolution, has been studied as a subsequent task of syntactic parsing. Most existing studies of semantic analysis have used newspaper corpora with manual annotation. However, there are sources other than newspapers, such as encyclopedias, diaries and nov- els each with diverse styles in each genre. There are linguistic phenomena that rarely appear in newspa- pers such as requests and honorific expressions. To deal with texts that include the above phenomena, it is essential to build an annotated corpus that includes diverse-domain documents. Web pages include var- ious genres and text styles such as news articles, en- cyclopedia articles, blog and business pages. Using web pages as the target documents of annotation, we build a Japanese annotated corpus that consists of various genres. We annotate predicate-argument structures and anaphoric relations as semantic relations. We il- lustrate these relations and annotations in Example (1)1. “A←rel:B” represents annotating B to A with relation rel. In the following examples, we some- times omit annotations that are not related to the dis- cussion. (1) a. ��� Taro-TOP ��� watch-ACC ���� ‘Taro bought a watch.’bought. (��� ← GA:��, WO:��) b. �� Little brother-DAT ��� it-ACC ���� gave ‘He gave it to his little brother.’ ⎛ ⎝ � ←NO:�� �� ←=:�� ��� ←GA:��, NI:��, WO:� ⎞ ⎠ Predicate-argument structures express the rela- tions between a predicate and its arguments. In Ex- ample (1a), the GA (nominative) case of ��� (bought) is �� (Taro) and the WO (accusative) case of ��� is �� (watch). In this example, there is a topic marker (�) which hides the case re- lation between �� and ���. Since the hidden actual case relation is GA, we annotate which the GA case of ��� is ��. Such disappearances 1In this paper, we use the following abbreviations: NOM (nominative), ABL(ablative), ACC (accusative), DAT (dative), ALL (allative), GEN (genitive), CMI (comitative), CNJ (con- junction), INS(instrumental) and TOP (topic marker). 535 of case markers occur also when a topic marker � …

Content truncated to 500 words