An Empirical Study of Span Representations in Argumentation Structure Parsing

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4691–4698 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 4691 An Empirical Study of Span Representations in Argumentation Structure Parsing Tatsuki Kuribayashi1, Hiroki Ouchi2,1, Naoya Inoue1,2, Paul Reisert2,1, Toshinori Miyoshi3, Jun Suzuki1,2, and Kentaro Inui1,2 1Tohoku University 2RIKEN Center for Advanced Intelligence Project (AIP) 3Research & Development Group, Hitachi, Ltd. {kuribayashi, naoya-i, jun.suzuki, inui} @ecei.tohoku.ac.jp {hiroki.ouchi, paul.reisert} @riken.jp toshinori.miyoshi.pd@hitachi.com Abstract For several natural language processing (NLP) tasks, span representation is attracting con- siderable attention as a promising new tech- nique; a common basis for an effective de- sign has been established. With such basis, exploring task-dependent extensions for argu- mentation structure parsing (ASP) becomes an interesting research direction. This study in- vestigates (i) span representation originally de- veloped for other NLP tasks and (ii) a sim- ple task-dependent extension for ASP. Our ex- tensive experiments and analysis show that these representations yield high performance for ASP and provide some challenging types of instances to be parsed. 1 Introduction Argumentation structure parsing (ASP) is the task of identifying argumentation structures in argu- mentative text. Figure 1 shows an example of an argumentative text and its structure. The structure forms a tree, the nodes of which are referred to as argumentative discourse units (ADUs) and the edges represent argumentative relations between the ADUs. ASP systems must identify such edges, edge labels (e.g., SUPPORT and ATTACK), and node labels (e.g., PREMISE, CLAIM, and MAJOR- CLAIM). A key to achieving high performance is feature representation design for segmental dis- course units (spans), such as ADUs. The aim of this study is to update the foundation of span rep- resentation design for ASP. Potash et al. (2017) introduced a model ex- ploiting neural network-based span representa- tion for ASP and reported state-of-the-art perfor- mance. Similarly, for other natural language pro- cessing (NLP) tasks, such as syntactic and seman- tic parsing, neural network-based span representa- tion design is attracting considerable attention as a promising new technique; some effective designs ADU1: In addition, I believe that city provides more work opportunities than the countryside. ADU2: There are not only more jobs, but they are also well-paid. ADU3: Of course living in a city is more expensive, ADU1 ADU2 ADU3 ADU4: but incomes are higher too. ADU4 ADU1 ADU2 ADU3 ADU4 (i) Current SOTA (ii) LSTM+dist (ELMo) Gold structure ADU1: In addition, I believe that city provides more work opportunities than the countryside. ADU2: There are not only more jobs, but they are also well-paid. ADU3: Of course living in a city is more expensive, ADU4: but incomes are higher too. support attack attack Premise Premise Claim Premise support attack attack ADU1: In addition, I believe that city provides more work opportunities than the countryside. ADU2: There are not only more jobs, but they are also well-paid. ADU3: Of course living in a city is more expensive, ADU4: but incomes are higher too. support attack attack Figure 1: Example of …

Content truncated to 500 words