Building a Macro Chinese Discourse Treebank

Building a Macro Chinese Discourse Treebank Xiaomin Chu, Feng Jiang, Sheng Xu, Qiaoming Zhu Natural Language Processing Lab Soochow University {xmchu, fjiang, sxu}@stu.suda.edu.cn, qmzhu@suda.edu.cn Abstract Discourse structure analysis is an important research topic in natural language processing. Discourse structure analysis not only helps to understand the discourse structure and semantics, but also provides strong support for deep applications of natural language processing, such as automatic summarization, statistical machine translation, question and answering, etc. At present, the analyses of discourse structure are mainly concentrated on the micro level, while the analyses on macro level are few. Therefore, this paper focuses on the construction of representation schema and corpus resources on the macro level of discourse structure. This paper puts forward a macro discourse structure framework and constructs the logical semantic structure and functional pragmatic structure respectively. On this basis, a macro Chinese discourse structure treebank is annotated, consisting of 147 Newswire articles. Preliminary experimental results show that the representation schema and corpus resource constructed in this paper can lay the foundation for further analysis of macro discourse structure. Keywords: discourse structure, macro discourse structure corpus, logical semantic structure, functional pragmatic structure 1. Introduction A discourse is not formed by independent and isolated dis- course units, but by related and structure units. The task of discourse analysis is to segment sentences into elementary discourse units(EDUs) and recognize the relations among them to form a complete discourse structure. Due to the se- mantic integrity of discourse units, discourse relations and their well-formed structure, discourse informations have been applied to many natural language processing applica- tions, such as information retrieval (Zou et al., 2014), auto- matic summarization (Ferreira et al., 2014; Cohan and Go- harian, 2017), question and answering (Sadek and Meziane, 2016) and statistical machine translation (Guzm´an et al., 2014). Previous research works have proven that discourse informations are beneficial to these NLP applications. The advent of large-scale collections of annotated data shifted the research community of natural language pro- cessing. These corpora have accelerated the development efforts and energized the research community. Generally speaking, there exist two hierarchical levels of discourse structures: micro level and macro level. At present, the analyses of discourse structure are mainly con- centrated on the micro level, that is, the relations and struc- tures between sentences or sentence groups. But the anal- yses on macro level are relatively few, that is, the relations and structures between paragraphs or documents. Through the above analysis, it is obvious to realize that de- veloping a macro discourse structure corpus is helpful to understand the overall discourse information and quite nec- essary for macro natural language processing tasks. 2. Macro Discourse Structure Framework The overall discourse structure is relevant to the discourse genre and discourse pattern. Thus discourse structures vary if the genres are different. For example, news articles are commonly described in “summary-story” structure, and academic papers are consist of “abstract, introduction, re- lated work, experimentation, conclusion”, while court doc- uments are recorded in the structure of “in what way, …

Content truncated to 500 words