Can Unsupervised Knowledge Transfer from Social Discussions Help Argument Mining?

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 7774 - 7786 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Can Unsupervised Knowledge Transfer from Social Discussions Help Argument Mining? Subhabrata Dutta∗ Jadavpur University subha0009@gmail.com Jeevesh Juneja∗ Delhi Technological University creativityinczenyoga@gmail.com Dipankar Das Jadavpur University dipankar.dipnil@gmail.com Tanmoy Chakraborty IIIT-Delhi tanmoy@iiitd.ac.in Abstract Identifying argument components from un- structured texts and predicting the relation- ships expressed among them are two primary steps of argument mining. The intrinsic com- plexity of these tasks demands powerful learn- ing models. While pretrained Transformer- based Language Models (LM) have been shown to provide state-of-the-art results over different NLP tasks, the scarcity of manu- ally annotated data and the highly domain- dependent nature of argumentation restrict the capabilities of such models. In this work, we propose a novel transfer learning strategy to overcome these challenges. We utilize argumentation-rich social discussions from the ChangeMyView subreddit as a source of unsupervised, argumentative discourse-aware knowledge by finetuning pretrained LMs on a selectively masked language modeling task. Furthermore, we introduce a novel prompt- based strategy for inter-component relation prediction that compliments our proposed fine- tuning method while leveraging on the dis- course context. Exhaustive experiments show the generalization capability of our method on these two tasks over within-domain as well as out-of-domain datasets, outperforming several existing and employed strong baselines.1 1 Introduction Computational argument mining from texts is the fine-grained process of understanding opinion dy- namics. In the most fundamental sense, argument understanding requires the identification of the opinions posed and justifications provided to sup- port or falsify them. Generally, automated argu- ment mining is a multi-stage pipeline identified with three general steps (Lippi and Torroni, 2015; Stab and Gurevych, 2017) – separating argumen- tative spans from non-argumentative ones, classi- ∗*Equal contribution 1We release all code, models and data used at https: //github.com/Jeevesh8/arg_mining Figure 1: Token-level claim (red) and premise (blue) annota- tion of a discussion thread formed by consecutive posts from two users. Second post quotes a span from the first (shown in italics). Highlighted regions signify component boundaries (to demarcate consecutive components of the same kind as in the fourth post). fying argument components, and inducing a struc- ture among them (support, attack, etc.). While different argumentation models define different taxonomies for argument components, popular ap- proaches broadly categorize them as ‘claims’ and ‘premises’ (Stab and Gurevych, 2017; Egawa et al., 2019; Mayer et al., 2020). As these components are not necessarily aligned to sentence-level segments and can be reflected within clausal levels, the task of argument component identification requires a token-level boundary detection of components and component type classification. Context of argumentation in online discus- sions. Online discussions originating from back- and-forth posts from users reflect a rich interaction of opinion dynamics on large scale. In Figure 1, we show a sample argument component annotation of consecutive posts from two users. The token-level granularity of components ensures that a single sentence may contain multiple components of the same (in 1st …

Content truncated to 500 words