Joint Syntacto-Discourse Parsing and the Syntacto-Discourse Treebank

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2117–2123 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Joint Syntacto-Discourse Parsing and the Syntacto-Discourse Treebank ∗ Kai Zhao† and Liang Huang School of Electrical Engineering and Computer Science Oregon State University Corvallis, Oregon, USA {kzhao.hf, liang.huang.sh}@gmail.com Abstract Discourse parsing has long been treated as a stand-alone problem independent from constituency or dependency parsing. Most attempts at this problem are pipelined rather than end-to-end, sophisticated, and not self-contained: they assume gold- standard text segmentations (Elementary Discourse Units), and use external parsers for syntactic features. In this paper we propose the first end-to-end discourse parser that jointly parses in both syntax and discourse levels, as well as the first syntacto-discourse treebank by integrating the Penn Treebank with the RST Tree- bank. Built upon our recent span-based constituency parser, this joint syntacto- discourse parser requires no preprocessing whatsoever (such as segmentation or fea- ture extraction), achieves the state-of-the- art end-to-end discourse parsing accuracy. 1 Introduction Distinguishing the semantic relations between segments in a document can be greatly beneficial to many high-level NLP tasks, such as summariza- tion (Louis et al., 2010; Yoshida et al., 2014), sen- timent analysis (Voll and Taboada, 2007; Soma- sundaran et al., 2009; Bhatia et al., 2015), ques- tion answering (Ferrucci et al., 2010; Jansen et al., 2014), and textual quality evaluation (Tetreault et al., 2013; Li and Jurafsky, 2016). There has been a variety of research on dis- course parsing (Marcu, 2000a; Soricut and Marcu, 2003; Pardo and Nunes, 2008; Hernault et al., ∗ The source code and the joint treebank are available at https://github.com/kaayy/josydipa. † Current address: Google Inc., New York, NY, USA. 2010; da Cunha et al., 2012; Joty et al., 2013; Joty and Moschitti, 2014; Feng and Hirst, 2014; Ji and Eisenstein, 2014; Li et al., 2014a,b; Heilman and Sagae, 2015; Wang et al., 2017). But most of them suffer from the following limitations: 1. pipelined rather than end-to-end: they as- sume pre-segmented discourse, and worse yet, use gold-standard segmentations, except Hernault et al. (2010); 2. not self-contained: they rely on external syn- tactic parsers and pretrained word vectors; 3. complicated: they design sophisticated fea- tures, including those from parse-trees. We argue for the first time that discourse parsing should be viewed as an extension of, and be per- formed in conjunction with, constituency parsing. We propose the first joint syntacto-discourse tree- bank, by unifying constituency and discourse tree representations. Based on this, we propose the first end-to-end incremental parser that jointly parses at both constituency and discourse levels. Our algo- rithm builds up on the span-based parser (Cross and Huang, 2016); it employs the strong general- ization power of bi-directional LSTMs, and parses efficiently and robustly with an extremely simple span-based feature set that does not use any tree structure information. We make the following contributions: 1. We develop a combined representation of constituency and discourse trees to facilitate parsing at both levels without explicit conver- sion mechanism. …

Content truncated to 500 words