Unifying Discourse Resources with Dependency Framework

Unifying Discourse Resources with Dependency Framework Yi Cheng1 Sujian Li1 Yueyuan Li2 1MOE Key Lab of Computational Linguistics, School of EECS, Peking University 2School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University {yicheng, lisujian}@pku.edu.cn, rowena lee@sjtu.edu.cn Abstract For text-level discourse analysis, there are various discourse schemes but relatively few labeled data, because discourse research is still immature and it is labor-intensive to annotate the inner logic of a text. In this paper, we attempt to unify multiple Chinese discourse corpora under different annotation schemes with discourse dependency framework by designing semi-automatic methods to convert them into dependency structures. We also implement several benchmark dependency parsers and research on how they can leverage the unified data to improve performance.1 1 Introduction Discourse parsing aims to construct the logical structure of a text and label relations between discourse units, as shown in Fig. 1 and Fig. 2. Various discourse corpora have been developed to promote the discourse parsing technique. There exist multiple discourse schemes such as Rhetorical Structure Theory (RST) (Carlson et al., 2001) and PDTB (Prasad et al., 2008), which act as guidelines of various discourse corpora. At the same time, text-level discourse annotation is complicated and laborious, so the scale of a single discourse corpora is often much smaller compared with other NLP tasks. One way to conquer data sparsity is to use different discourse corpora through multi-task learning. One way to conquer the data-sparsity problem is to multi-task learning multiple corpora simultaneously. Prior efforts on that mainly focus on discourse relation classification between two text spans, without considering the whole discourse structure (Liu et al., 2016; Li et al., 2017), since one principle of multi- task learning is that the tasks should be closely related, but the discourse structures of different corpora vary a lot, e.g. shallow predicate-arguments structure (e.g. the relations in Fig. 1) and deep tree structure (e.g. the CDTB tree in Fig. 2). In this work, we explore another possible way to simultaneously leverage different corpora: we unify the existing discourse corpora under one same framework to form a much larger dataset. Our choice for this unified framework is dependency discourse structure (DDS), where elementary discourse units (EDUs) are directly related to each other without consideration of intermediate text spans. Because as Morey et al. (2018) have argued, DDS is a very general discourse representation framework, and they adopt a dependency perspective to evaluate the English discourse corpora RST-DT (Carlson et al., 2001) and the related parsing techniques. Our work explore the feasibility of unifying discourse resources under dependency framework with consideration of the following two questions: (i) How to convert other discourse structures into DDS? (ii) How to make the best use of the unified data to improve discourse parsing techniques? Oriented by the questions above, we unify three Chinese discourse corpora under dependency frame- work: HIT-CDTB (Zhang et al., 2014), SU-CDTB (Li et al., 2014b) and Sci-CDTB (Cheng and Li, 2019). HIT-CDTB adopts the predicate-argument structure similar to PDTB, with a connective …

Content truncated to 500 words