Uncovering Implicit Inferences for Improved Relational Argument Mining

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2484–2495 May 2-6, 2023 ©2023 Association for Computational Linguistics Uncovering Implicit Inferences for Improved Relational Argument Mining Ameer Saadat-Yazdi School of Informatics University of Edinburgh ameer.saadat@ed.ac.uk Jeff Z. Pan School of Informatics University of Edinburgh j.z.pan@ed.ac.uk Nadin Kökciyan School of Informatics University of Edinburgh nadin.kokciyan@ed.ac.uk Abstract Argument mining seeks to extract arguments and their structure from unstructured texts. Identifying relations (such as attack, support, and neutral) between argumentative units is a challenging task because two units may be related to each other via implicit infer- ences. These inferences often rely on ex- ternal commonsense knowledge to discover how one argumentative unit relates to another. State-of-the-art methods, however, rely on pre- defined knowledge graphs, and thus might not cover target pairs of argumentative units well. We introduce a new generative ap- proach to finding inference chains that con- nect these pairs by making use of the Com- monsense Transformer (COMET). We evalu- ate our approach on three datasets for both the two-label (attack/support) and three-label (at- tack/support/neutral) tasks. Our approach sig- nificantly outperforms the state-of-the-art, by 2- 5% in F1 score, on two out of the three datasets with minor improvements on the remaining one. 1 Introduction In argumentation, such as in a debate, it is impor- tant to understand the key arguments put forward and how the arguments relate to each other based on a specific context. Ideally, we would like to empower systems to extract arguments and the re- lations between arguments automatically. Such systems could be used to aggregate opinions (Co- carascu et al., 2019), participate in debates (Slonim et al., 2021), or they could assist humans in making informed decisions by considering different points of view. Toulmin (1958) defines an argument as a claim backed by one or more grounds (or premises) based on a warrant (i.e., the inference link between the grounds and the claim). For example, consider the following two sentences: Drunk driving hurts inno- cent people. Therefore, drunk driving is wrong. Based on Toulmin’s model, we have the claim drunk driving is wrong, supported on the grounds that drunk driving hurts innocent people. In what follows, we will call the claim and grounds argu- mentative units (AUs). To justify the claim, the au- thor relies on the reader’s ability to uncover the im- plicit warrant that hurting innocent people is wrong which connects the grounds to the claim. While Toulmin’s warrants model only arguments in which the units are connected by support relations, we ex- tend the definition to also allow for attacks and neu- tral relations. We do this to enable us to apply the notion of warrants to the relational argument min- ing task, as proposed by Carstens and Toni (2015), to better identify attack/support/neutral relations between pairs of AUs. Finding the correct warrant is a challenging task that requires deep reasoning, which pre-trained lan- guage models struggle with (Helwe et al., 2021). However, by iteratively …

Content truncated to 500 words