Split or Merge: Which is Better for Unsupervised RST Parsing?

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 5797–5802, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 5797 Split or Merge: Which is Better for Unsupervised RST Parsing? Naoki Kobayashi†, Tsutomu Hirao‡, Kengo Nakamura‡, Hidetaka Kamigaito†, Manabu Okumura†, Masaaki Nagata‡ † Institute of Innovative Research, Tokyo Institute of Technology, ‡ NTT Communication Science Laboratories, NTT Corporation {kobayasi, kamigaito}@lr.pi.titech.ac.jp, oku@pi.titech.ac.jp {tsutomu.hirao.kp, kengo.nakamura.dx, masaaki.nagata.et}@hco.ntt.co.jp Abstract Rhetorical Structure Theory (RST) parsing is crucial for many downstream NLP tasks that require a discourse structure for a text. Most of the previous RST parsers have been based on supervised learning approaches. That is, they require an annotated corpus of sufficient size and quality, and heavily rely on the language and domain dependent corpus. In this paper, we present two language-independent unsu- pervised RST parsing methods based on dy- namic programming. The first one builds the optimal tree in terms of a dissimilarity score function that is defined for splitting a text span into smaller ones. The second builds the opti- mal tree in terms of a similarity score function that is defined for merging two adjacent spans into a large one. Experimental results on En- glish and German RST treebanks showed that our parser based on span merging achieved the best score, around 0.8 F1 score, which is close to the scores of the previous supervised parsers. 1 Introduction Rhetorical Structure Theory (RST) (Mann and Thompson, 1988) is one of the theories that are most widely utilized for representing a discourse structure of a text in downstream NLP applica- tions, such as automatic summarization (Marcu, 1998; Hirao et al., 2013), sentiment analysis (Bha- tia et al., 2015), and text categorization (Ji and Smith, 2017). RST represents a text as a kind of constituent tree, whose leaves are Elementary Dis- course Units (EDUs), clause-like units, and whose non-terminal nodes cover text spans consisting of a sequence of EDUs or a singleton EDU. The label of a non-terminal node represents the attribution of a text span, nucleus or satellite. A discourse re- lation is also assigned between two adjacent non- terminal nodes. Since the RST tree can be regarded as a stan- dard constituent (phrase structure) tree, syntactic parsing models for constituency parsing can also be successfully applied to RST parsing. In most cases, RST parsers have been developed on the basis of supervised learning algorithms, which re- quire a high quality annotated corpus of sufficient size. As a result, research on RST parsing has fo- cused on English, with the largest annotated cor- pus being the RST Discourse Treebank (RST-DT) (Carlson et al., 2001). These supervised RST pars- ing methods might not be applied to languages with only a small-size corpus. This paper presents two types of language inde- pendent unsupervised RST parsing methods based on the CKY-like dynamic programming-based ap- proach. One method builds the most likely parse tree in terms of a dissimilarity score defined for …

Content truncated to 500 words