AMPERSAND: Argument Mining for PERSuAsive oNline Discussions

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 2933–2943, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 2933 AMPERSAND: Argument Mining for PERSuAsive oNline Discussions Tuhin Chakrabarty1, Christopher Hidey1, Smaranda Muresan1,2 , Kathleen Mckeown1 and Alyssa Hwang1 1Department of Computer Science, Columbia University 2Data Science Institute, Columbia University {tuhin.chakrabarty, smara, a.hwang}@columbia.edu {chidey, kathy}@cs.columbia.edu Abstract Argumentation is a type of discourse where speakers try to persuade their audience about the reasonableness of a claim by presenting supportive arguments. Most work in argument mining has focused on modeling arguments in monologues. We propose a computational model for argument mining in online persua- sive discussion forums that brings together the micro-level (argument as product) and macro- level (argument as process) models of argu- mentation. Fundamentally, this approach re- lies on identifying relations between compo- nents of arguments in a discussion thread. Our approach for relation prediction uses contex- tual information in terms of fine-tuning a pre- trained language model and leveraging dis- course relations based on Rhetorical Struc- ture Theory. We additionally propose a can- didate selection method to automatically pre- dict what parts of one’s argument will be tar- geted by other participants in the discussion. Our models obtain significant improvements compared to recent state-of-the-art approaches using pointer networks and a pre-trained lan- guage model. 1 Introduction Argument mining is a field of corpus-based discourse analysis that involves the automatic identification of argumentative structures in text. Most current studies have focused on mono- logues or micro-level models of argument that aim to identify the structure of a single argu- ment by identifying the argument components (classes such as “claim” and “premise”) and re- lations between them (“support” or “attack”) (So- masundaran et al., 2007; Stab and Gurevych, 2014; Swanson et al., 2015; Feng and Hirst, 2011; Habernal and Gurevych, 2017; Peldszus and Stede, 2015). Macro-level models (or dialogical models) and rhetorical models which focus on the Figure 1: Annotated Discussion Thread (C: Claim, P: Premise, MC: Main Claim) process of argument in a dialogue (Bentahar et al., 2010) have received less attention. We propose a novel approach to automatically identify the argument structure in persuasive di- alogues that brings together the micro-level and the macro-level models of argumentation. Our ap- proach identifies argument components in a full discussion thread and two kinds of argument re- lations: inter-turn relations (argumentative rela- tions to support or attack the other person’s ar- gument) and intra-turn relations (to support one’s claim). Figure 1 shows a thread structure consist- ing of multiple posts with argumentative compo- nents (main claims, claims or premises) and both intra- and inter-turn relations. We focus here on the relation identification task (i.e, predicting the existence of a relation between two argument com- ponents). To learn to predict relations, we annotate argu- mentative relations in threads from the Change 2934 My View (CMV) subreddit. In addition to the CMV dataset we also introduce two novel …

Content truncated to 500 words