Employing Text Matching Network to Recognise Nuclearity in Chinese Discourse

Proceedings of the 27th International Conference on Computational Linguistics, pages 525–535 Santa Fe, New Mexico, USA, August 20-26, 2018. 525 Employing Text Matching Network to Recognise Nuclearity in Chinese Discourse Sheng Xu1, Peifeng Li1, Guodong Zhou1, and Qiaoming Zhu12 1School of Computer Science and Technology, Soochow University, China 2Institute of Artificial Intelligence, Soochow University, China sxu@stu.suda.edu.cn;{pfli, gdzhou, qmzhu}@suda.edu.cn Abstract The task of nuclearity recognition in Chinese discourse remains challenging due to the demand for more deep semantic information. In this paper, we propose a novel text matching network (TMN) that encodes the discourse units and the paragraphs by combining Bi-LSTM and CNN to capture both global dependency information and local n-gram information. Moreover, it in- troduces three components of text matching, the Cosine, Bilinear and Single Layer Network, to incorporate various similarities and interactions among the discourse units. Experimental results on the Chinese Discourse TreeBank show that our proposed TMN model significantly outper- forms various strong baselines in both micro-F1 and macro-F1. 1 Introduction During the past few years, the focus of Natural Language Understanding (NLU) has shifted from the word/sentence level to the discourse level. A challenging task in NLU is discourse parsing, which in- volves analysing the relations between discourse units and building the document structure. As one of the most influential discourse theories, Rhetorical Structure Theory (RST) (Mann and Thompson, 1988) defines a document as a collection of Elementary Discourse Units (EDUs) with semantic connections and combines adjacent EDUs with rhetorical relations in a hierarchical way to represent an entire doc- ument as a discourse tree. As a critical subtask in discourse parsing, nuclearity recognition involves identifying the nuclearity between the discourse units and thus being able to extract the main information in a document. Accord- ing to RST, a discourse relation can be divided into mononuclear and multinuclear. A mononuclear re- lation holds a nucleus and a satellite, where the nucleus expresses the main textual information and the satellite offers additional information about the nucleus (Stede, 2008), while a multinuclear relation holds two or more discourse units, which are all nuclei. Therefore, three types of nuclearity exist: Nu- cleus-Satellite if the left subtree is the nucleus and the right subtree is the satellite, Satellite-Nucleus if the order of the satellite and nucleus is inverted, and Nucleus-Nucleus for multinuclear relations. Nuclearity recognition is helpful in detecting discourse relations (Iruskieta et al., 2014) and extracting the main content of a document, and it is widely used in various NLP tasks, including automatic sum- marisation (Louis et al., 2010; Marcu, 2000), question answering (Verberne et al., 2007) and information extraction (Zou et al., 2014). Consider the following document as an example: Example 1: 中国机电产品进出口贸易继续增加 a，占总进出口的比重继续上升 b。其中，出口 五十七点九亿美元 c，占总出口的百分之三十二点五 d；进口八十五点二亿美元 e，占总进口的 百分之四十六点四 f，均比去年同期有所上升 g。The import and export trade of China’s mechanical and electronic products continues to increase a, and its proportion of the total imports and exports also continues to rise b. Among them, the exports amounted to 5.79 billion dollars c, accounting for 32.5 percent of the total exports …

Content truncated to 500 words