Transferring Confluent Knowledge to Argument Mining

Proceedings of the 29th International Conference on Computational Linguistics, pages 6859–6874 October 12–17, 2022. 6859 Transferring Confluent Knowledge to Argument Mining João A. Rodrigues and António Branco University of Lisbon NLX – Natural Language and Speech Group, Department of Informatics Faculdade de Ciências, Campo Grande, 1749-016 Lisboa, Portugal {jarodrigues,ambranco}@fc.ul.pt Abstract Relevant to all application domains where it is important to get at the reasons underlying senti- ments and decisions, argument mining seeks to obtain structured arguments from unstructured text and has been addressed by approaches typ- ically involving some feature and/or neural ar- chitecture engineering. By adopting a transfer learning methodology, and by means of a systematic study with a wide range of knowledge sources promisingly suit- able to leverage argument mining, the aim of this paper is to empirically assess the potential of transferring such knowledge learned with confluent tasks. By adopting a lean approach that dispenses with heavier feature and model engineering, this study permitted both to gain novel empiri- cally based insights into the argument mining task and to establish new state of the art levels of performance for its three main sub-tasks, viz. identification of argument components, classi- fication of the components, and determination of the relation among them. 1 Introduction Argument mining is a Natural Language Process- ing (NLP) task consisting in taking unstructured text as input and returning it annotated such that each portion occurring in it that is an argument is properly delimited and analysed (Schneider et al., 2013; Peldszus and Stede, 2013; Lippi and Torroni, 2016; Habernal and Gurevych, 2017; Wachsmuth et al., 2017; Stede and Schneider, 2018; Lawrence and Reed, 2020). Argument mining relates to the high-level human capacity of reasoning (Walton et al., 2005), it is at the core of social interaction concerned with persuasion (Mercier and Sperber, 2017), and it is of utmost importance to enhance applications across different domains that aim at enhancing their services beyond mere sentiment analysis, on the basis of the reasons uncovered for the associated sentiments and decisions (Habernal et al., 2014). Argument mining has been decomposed into a number of sub-tasks. While the number and profiling of these tasks depends on the theo- retical approach adopted to analyse arguments (Van Eemeren et al., 2019), they typically involve some sort of delimitation of the text segments con- veying argument components, the classification of the roles of these components (e.g. premises, con- clusions, etc.), and the classification of the type of relation among those components (e.g. support, attack, etc.) (Lawrence and Reed, 2020). These sub-tasks and their eventual pipeline in argument mining have been addressed by means of supervised deep learning approaches that in- volve some degree of neural architecture engineer- ing (Eger et al., 2017; Potash et al., 2017; Nguyen and Litman, 2016) a.o. Recently, first attempts to approach argument mining with Transformers have been reported in the literature (Wang et al., 2020; Rodrigues et al., 2020a) a.o., tough at an ex- ploratory level that leaves much of its strength still untapped. This has …

Content truncated to 500 words