Chinese Discourse Parsing: Model and Evaluation

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 1019–1024 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 1019 Chinese Discourse Parsing: Model and Evaluation Chuan-An Lin1, Shyh-Shiun Hung1, Hen-Hsen Huang2, Hsin-Hsi Chen1 1Department of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan 2Department of Computer Science, National Chengchi University, Taipei, Taiwan 3MOST Joint Research Center for AI Technology and All Vista Healthcare, Taipei, Taiwan calin@nlg.csie.ntu.edu.tw, shhung@nlg.csie.ntu.edu.tw, hhhuang@nccu.edu.tw, hhchen@ntu.edu.tw Abstract Chinese discourse parsing, which aims to identify the hierarchical relationships of Chinese elementary discourse units, has not yet a consistent evaluation metric. Although Parseval is commonly used, variations of evaluation differ from three aspects: micro vs. macro F1 scores, binary vs. multiway ground truth, and left-heavy vs. right-heavy binarization. In this paper, we first propose a neural network model that unifies a pre-trained transformer and CKY-like algorithm, and then compare it with the previous models with different evaluation scenarios. The experimental results show that our model outperforms the previous systems. We conclude that (1) the pre-trained context embedding provides effective solutions to deal with implicit semantics in Chinese texts, and (2) using multiway ground truth is helpful since different binarization approaches lead to significant differences in performance. Keywords: Discourse parsing, CKY, Shift-reduce, PARSEVAL 1. Introduction As mentioned by Rhetorical Structure Theory (Mann and Thompson, 1988), a discourse is composed of elementary discourse units (EDUs), which can be formed into a hierar- chical structure by relating each other with discourse rela- tions. This kind of discourse parsing tree provides a deep understanding of an article and benefits downstream NLP tasks. The majority of Chinese discourse relation is “implicit”, lacking obvious lexical cues to discriminate the relation type (Li et al., 2014b). It makes the task of Chinese dis- course parsing more challenging as the parser has to catch the implicit meaning from the text instead of relying on lex- ical information. According to the Connective-Driven Dependency Tree (CDT) scheme (Li et al., 2014b), there are four subtasks in Chinese discourse parsing, including EDU segmentation, tree structure construction, relation sense labeling, and re- lation center labeling. Figure 1 illustrates an example of Chinese discourse parsing tree. Note that a coordination relation may have more than two arguments while the other kinds of relations are always binary. Details of Chinese dis- course parsing can be found in (Li et al., 2014b). The CDT annotates the hierarchical discourse structure of a given ar- ticle, which is different from the PDTB-style scheme pro- posed by (Zhou and Xue, 2012). Although the annotation of the Chinese Discourse Tree- Bank (CDTB) is well-defined, the evaluation is divergent. Generally, PARSEVAL (Black et al., 1991) is used to eval- uate the quality of a predicted parsing tree. For a golden standard discourse tree, we have a set of non-leaf nodes N = {n1, n2, ..., nk}. We also have a set of text spans T = {t1, t2, ..., tk} where ni dominates ti for all i (e.g., …

Content truncated to 500 words