Neural End-to-End Learning for Computational Argumentation Mining

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 11–22 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1002 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 11–22 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1002 Neural End-to-End Learning for Computational Argumentation Mining Steffen Eger†‡, Johannes Daxenberger†, Iryna Gurevych†‡ †Ubiquitous Knowledge Processing Lab (UKP-TUDA) Department of Computer Science, Technische Universitt Darmstadt ‡Ubiquitous Knowledge Processing Lab (UKP-DIPF) German Institute for Educational Research and Educational Information http://www.ukp.tu-darmstadt.de Abstract We investigate neural techniques for end- to-end computational argumentation min- ing (AM). We frame AM both as a token- based dependency parsing and as a token- based sequence tagging problem, includ- ing a multi-task learning setup. Contrary to models that operate on the argument component level, we find that framing AM as dependency parsing leads to subpar per- formance results. In contrast, less com- plex (local) tagging models based on BiL- STMs perform robustly across classifica- tion scenarios, being able to catch long- range dependencies inherent to the AM problem. Moreover, we find that jointly learning ‘natural’ subtasks, in a multi-task learning setup, improves performance. 1 Introduction Computational argumentation mining (AM) deals with finding argumentation structures in text. This involves several subtasks, such as: (a) separating argumentative units from non-argumentative units, also called ‘component segmentation’; (b) classi- fying argument components into classes such as “Premise” or “Claim”; (c) finding relations be- tween argument components; (d) classifying rela- tions into classes such as “Support” or “Attack” (Persing and Ng, 2016; Stab and Gurevych, 2017). Thus, AM would have to detect claims and premises (reasons) in texts such as the following, where premise P supports claim C: Since it killed many marine livesP , ::::::: tourism ::: has:::::::::: threatened:::::: natureC . Argument structures in real texts are typically much more complex, cf. Figure 1. While different research has addressed different subsets of the AM problem (see below), the ul- timate goal is to solve all of them, starting from unannotated plain text. Two recent approaches to this end-to-end learning scenario are Persing and Ng (2016) and Stab and Gurevych (2017). Both solve the end-to-end task by first training indepen- dent models for each subtask and then defining an integer linear programming (ILP) model that en- codes global constraints such as that each premise has a parent, etc. Besides their pipeline architec- ture the approaches also have in common that they heavily rely on hand-crafted features. Hand-crafted features pose a problem because AM is to some degree an “arbitrary” problem in that the notion of “argument” critically relies on the underlying argumentation theory (Reed et al., 2008; Biran and Rambow, 2011; Habernal and Gurevych, 2015; Stab and Gurevych, 2017). Ac- cordingly, datasets typically differ with respect to their annotation of (often rather complex) argu- ment structure. Thus, feature sets would have to be manually adapted to and designed for each new sample of data, a challenging task. …

Content truncated to 500 words