Chinese Paragraph-level Discourse Parsing with Global Backward and Local Reverse Reading

Proceedings of the 28th International Conference on Computational Linguistics, pages 5749–5759 Barcelona, Spain (Online), December 8-13, 2020 5749 Chinese Paragraph-level Discourse Parsing with Global Backward and Local Reverse Reading Feng Jiang, Xiaomin Chu∗, Peifeng Li, Fang Kong, Qiaoming Zhu School of Computer Science and Technology, Soochow University, China fjiang@stu.suda.edu.cn {xmchu, pfli, kongfang, qmzhu}@suda.edu.cn Abstract Discourse structure tree construction is the fundamental task of discourse parsing and most pre- vious work focused on English. Due to the cultural and linguistic differences, existing successful methods on English discourse parsing cannot be transformed into Chinese directly, especially in paragraph level suffering from longer discourse units and fewer explicit connectives. To allevi- ate the above issues, we propose two reading modes, i.e., the global backward reading and the local reverse reading, to construct Chinese paragraph level discourse trees. The former processes discourse units from the end to the beginning in a document to utilize the left-branching bias of discourse structure in Chinese, while the latter reverses the position of paragraphs in a discourse unit to enhance the differentiation of coherence between adjacent discourse units. The experi- mental results on Chinese MCDTB demonstrate that our model outperforms all strong baselines. 1 Introduction Discourse parsing aims to study the internal structure of texts and understand the semantic relationship between various kinds of text units, such as clauses, sentences, sentence groups, paragraphs, and chap- ters. Since it is fundamental to understanding the overall text semantics, discourse parsing has been widely applied to various Natural Language Processing (NLP) applications, such as sentiment analysis (Bhatia et al., 2015), question answering (Sadek and Meziane, 2016), and summarization (Cohan and Goharian, 2018; Xu et al., 2020). As one of the most influential theories in discourse parsing, Rhetorical Structure Theory (RST) (Mann and Thompson, 1987) represents text as a hierarchical structure known as a discourse tree. In the liter- ature, various RST-style corpora have been built, such as RST Discourse Treebank (RST-DT) (Carlson et al., 2003) and Macro Chinese Discourse Treebank (MCDTB) (Jiang et al., 2018b). In RST-style dis- course parsing, the parser first identifies whether there is a rhetorical relationship between discourse units to construct a naked tree and then recognizes the nuclearity and relation labels for each relationship, as shown in Figure 1. According to the granularity of the leaf nodes, the discourse tree is divided into three levels: clause level, sentence level and paragraph level (Kobayashi et al., 2020). This paper focuses on constructing paragraph-level Chinese discourse trees where the leaf node is a paragraph. We call a leaf node that contains only one paragraph as PDU (Paragraph-level Discourse Unit) to distinguish the ele- mentary discourse unit (EDU) at the clause level. It is more important for downstream tasks as the upper part of a complete discourse tree. In English, there is a series of successes (Feng and Hirst, 2014; Ji and Eisenstein, 2014; Wang et al., 2017) in RST-DT (Carlson et al., 2003), especially Lin et al. (2019) and Liu et al. (2019) got excellent performance in clause-level …

Content truncated to 500 words