End-to-End Argument Mining as Biaffine Dependency Parsing

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 669–678 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 669 End-to-End Argument Mining as Biaffine Dependency Parsing Yuxiao Ye University of Cambridge Cambridge, UK yy477@cl.cam.ac.uk Simone Teufel University of Cambridge Cambridge, UK sht25@cl.cam.ac.uk Abstract Non-neural approaches to argument mining (AM) are often pipelined and require heavy feature-engineering. In this paper, we propose a neural end-to-end approach to AM which is based on dependency parsing, in contrast to the current state-of-the-art which relies on rela- tion extraction. Our biaffine AM dependency parser significantly outperforms the state-of- the-art, performing at F1 = 73.5% for com- ponent identification and F1 = 46.4% for re- lation identification. One of the advantages of treating AM as biaffine dependency parsing is the simple neural architecture that results. The idea of treating AM as dependency parsing is not new, but has previously been abandoned as it was lagging far behind the state-of-the-art. In a thorough analysis, we investigate the fac- tors that contribute to the success of our model: the biaffine model itself, our representation for the dependency structure of arguments, differ- ent encoders in the biaffine model, and syntac- tic information additionally fed to the model. Our work demonstrates that dependency pars- ing for AM, an overlooked idea from the past, deserves more attention in the future. 1 Introduction People often hold different opinions about the same thing. One very common way to express one’s opin- ions is to construct an argument (Mercier and Sper- ber, 2011; Van Eemeren and Henkemans, 2016). To help people efficiently understand different opin- ions and their reasoning embedded in arguments, it is necessary to develop systems that can automati- cally analyse structures of arguments. An emerging research field called Argumentation Mining or Ar- gument Mining (AM) (Peldszus and Stede, 2013; Green et al., 2014) addresses this problem. To analyse the structure of arguments, AM typically proposes four subtasks: 1) component segmentation, i.e., cutting a raw sequence into text spans that are either argumentative or non- argumentative segments (only argumentative seg- ments are called argument components); 2) com- ponent classification, i.e., labelling each argument components with a tag in a pre-defined scheme, (e.g., “PREMISE” or “CLAIM”); 3) relation detec- tion, i.e., deciding if two argument components are directly related; and 4) relation classification, i.e., categorizing a detected relation into a class in a pre- defined scheme, (e.g., “ATTACK” or “SUPPORT”) (Persing and Ng, 2016; Eger et al., 2017; Haber- nal and Gurevych, 2017; Stab and Gurevych, 2017; Lawrence and Reed, 2020). The first two subtasks are often referred together as component identifica- tion, and the last two as relation identification. An example of the argument structure is illustrated in Figure 1. We see that the “rain” component acts as the premise, supporting the claim of “beautiful”. Some of the AM approaches that try to solve all four subtasks use a pipelined architecture. Inde- pendent models are first trained for each subtask; the final …

Content truncated to 500 words