Decompositional Argument Mining: A General Purpose Approach for Argument Graph Construction

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 516–526 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 516 Decompositional Argument Mining: A General Purpose Approach for Argument Graph Construction Debela Gemechu and Chris Reed Centre for Argument Technology, School of Science & Engineering University of Dundee, Dundee, UK d.t.gemechu@dundee.ac.uk, c.a.reed@dundee.ac.uk Abstract This work presents an approach decomposing propositions into four functional components and identify the patterns linking those com- ponents to determine argument structure. The entities addressed by a proposition are target concepts and the features selected to make a point about the target concepts are aspects. A line of reasoning is followed by providing evidence for the points made about the tar- get concepts via aspects. Opinions on target concepts and opinions on aspects are used to support or attack the ideas expressed by target concepts and aspects. The relations between aspects, target concepts, opinions on target concepts and aspects are used to infer the ar- gument relations. Propositions are connected iteratively to form a graph structure. The ap- proach is generic in that it is not tuned for a specific corpus and evaluated on three differ- ent corpora from the literature: AAEC, AMT, US2016G1tv and achieved an F score of 0.79, 0.77 and 0.64, respectively. 1 Introduction Argument mining is the process of identifying ar- gumentative structure contained within a text. It involves segmenting arguments into elementary discourse units (EDUs), distinguishing argumen- tative units from non-argumentative units, classi- fying argument components into classes such as premise and claim, identifying and labeling argu- ment relations between the components, and iden- tifying argument schemes. We are here aimed at mining argument structure from text segmented into EDUs (or, more precisely for argument min- ing, Argumentative Discourse Units, ADUs (Peld- szus and Stede, 2015)). Several argument mining approaches use fea- tures identified from individual EDUs and apply classifiers (Moens et al., 2007); others use features that span EDUs and apply dependency parsing (Muller et al., 2012), similarity (Lawrence et al., 2014), linguistic indicators (Villalba and Saint- Dizier, 2012) and their combinations (Lawrence and Reed, 2015). Recently, a neural end-to-end method for argument mining shows that depen- dency parsing outperforms an EDU-level classifier (Eger et al., 2017). Stab and Gurevych (2014b) use both EDU-level and cross-EDU features to im- prove performance. The EDU-spanning features used by these latter approaches include syntac- tic dependency and lexical overlap between the EDUs. For instance, Eger et al. (2017) applied token level syntactic dependency to learn the re- lations between EDUs. Even though cross-EDU tokens are used for argument mining, the nature of such tokens is not studied well. Following the same line of reasoning, similar- ity approaches use EDU level similarity to deter- mine argument structure. Lawrence et al. (2014) use Latent Dirichlet Allocation (LDA) topic mod- eling; Lawrence and Reed (2015) use WordNet1 Synset hierarchy to determine similarity between propositions. Such approaches start from a con- clusion and determine the most related proposi- tion …

Content truncated to 500 words