A Top-down Neural Architecture towards Text-level Parsing of Discourse Rhetorical Structure

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6386–6395 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 6386 A Top-Down Neural Architecture towards Text-Level Parsing of Discourse Rhetorical Structure Longyin Zhang1,2, Yuqing Xing1,2, Fang Kong1,2⇤, Peifeng Li1,2, Guodong Zhou1,2 1. Institute of Artificial Intelligence, Soochow University, China 2. School of Computer Science and Technology, Soochow University, China {lyzhang9,yqxing}@stu.suda.edu.cn {kongfang,pfli,gdzhou}@suda.edu.cn Abstract Due to its great importance in deep natural lan- guage understanding and various down-stream applications, text-level parsing of discourse rhetorical structure (DRS) has been drawing more and more attention in recent years. How- ever, all the previous studies on text-level dis- course parsing adopt bottom-up approaches, which much limit the DRS determination on local information and fail to well benefit from global information of the overall discourse. In this paper, we justify from both computa- tional and perceptive points-of-view that the top-down architecture is more suitable for text- level DRS parsing. On the basis, we propose a top-down neural architecture toward text-level DRS parsing. In particular, we cast discourse parsing as a recursive split point ranking task, where a split point is classified to different lev- els according to its rank and the elementary discourse units (EDUs) associated with it are arranged accordingly. In this way, we can determine the complete DRS as a hierarchi- cal tree structure via an encoder-decoder with an internal stack. Experimentation on both the English RST-DT corpus and the Chinese CDTB corpus shows the great effectiveness of our proposed top-down approach towards text- level DRS parsing. 1 Introduction Text-level parsing of discourse rhetorical structure (DRS) aims to identify the overall discourse struc- ture and the rhetorical relations between discourse units in a text. As a fundamental research topic in natural language processing, text-level DRS pars- ing plays an important role in text understanding and can benefit various down-stream applications, such as document summarization (Goyal and Eisen- stein, 2016), sentiment analysis (Choi et al., 2016), text categorization (Ji and Smith, 2017), pronoun ⇤Corresponding author Commentary (SN) Overall-branch (SN) Coordinating (NN) e7 e6 Coordinating (NS) e5 e4 Purpose (SN) Coordinating (NN) e3 e2 e1 e1: 西œˆLËËÔÅ⇤t·7”Ñ，/ Bank of Tibetan actively readjusts credit structure e2: Ân›úg⇢生ßIÕπß⇢Ñïe，/ Ensuring the investment of key industries such as husbandry produc- tion e3: †'˘Â⇢、˝ê、§⇢、⇢·I˙æÑc8 D—õîœ⇥/ Increase the normal supply of funds for industrial, energy, transportation, communications e4: ªt∞û7>A€π€一øC，/ Last year, the newly increased loan was 1.441 billion yuan e5: ‘⌦tû†kø⇢C⇥/ an increase of more than 800 million yuan compared to the previous year. e6: úg⇢生ß7>（⇧Ïv+7> ‘⌦t∞û€ π køC；/ The loans (including aid the poor loan) for agricultural and livestock production newly increased by 438 million yuan compared to the previous year e7: aG企⇢7>ûE:~⌃KmA一πk ⇥/ The increase in loans to township enterprises was 61.83% Figure 1: An example for DRS parsing, where the text consists of 3 sentences containing 7 EDUs. resolution (Sheng et al., 2017) and event temporal relation identification (Dai et al., 2019). According to Rhetorical Structure The- ory (RST) (Mann and Thompson, 1988), a …

Content truncated to 500 words