Leveraging Topic Relatedness for Argument Persuasion

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4401–4407 August 1–6, 2021. ©2021 Association for Computational Linguistics 4401 Leveraging Topic Relatedness for Argument Persuasion Xinran Zhao1, Esin Durmus2, Hongming Zhang1, and Claire Cardie2 1The Hong Kong University of Science and Technology 2Cornell University xzhaoar@connect.ust.hk, ed459@cornell.edu hzhangal@cse.ust.hk, cardie@cs.cornell.edu Abstract Argumentation exposes individuals to conflict- ing viewpoints and can help them make more informed decisions based on the pros and cons of a particular issue. While recent studies of ar- gumentation in Natural Language Processing have mainly focused on understanding the ef- fect of various factors of persuasion (i.e. the source, audience, and language style), the im- pact of exploiting the relationships among con- troversial topics when predicting argument per- suasiveness remains under-explored. In this paper, we model the relatedness among con- troversial topics utilizing an embedding-based method based on individuals’ stances on the topics. We then leverage these topic embed- ding features and incorporate topic semantics features extracted from the arguments along with the previously studied factors of persua- sion. We show that incorporating both types of topic relatedness features explicitly leads to significant improvement in predicting persua- siveness and also helps enhance generalization to rare topics, in a few-shot setting. 1 Introduction Emergence of social media and online argumenta- tive forums provide users with a platform to gain information, express, and form opinions on a di- verse set of controversial topics (i.e. issues). The increasing importance of these online platforms has motivated NLP researchers to use these platforms as one of the main domains to study the impor- tant factors of persuasion. In particular, prior work has shown that characteristics of the speaker (i.e. source), prior beliefs of the audience (Lukin et al., 2017; Durmus and Cardie, 2018), and language style (Feng and Hirst, 2011; Tan et al., 2016) are important factors in determining persuasiveness of the arguments in online argumentation platforms. Although there has been evidence in previous stud- ies of Social Sciences that people’s perceptions on a particular controversial topic may be related to their perceptions on other controversial topics (Judd and Krosnick, 1989; Sapra, 2012), the impact of exploiting this relationship among controversial topics are under-explored in NLP studies of persua- sion. In this paper, we explicitly study the effect of incorporating topic relatedness among controver- sial topics in predicting argument persuasiveness. To study the impact of involving topic related- ness in argument persuasion, we define two types of features: (1) topic embedding features and (2) topic semantics features. Prior work has shown that topic is an important factor (Das et al., 2016) to determine whether an emotional vs. a logical argument will be received positively by the audi- ence. We hypothesize that encoding underlying relationship among topics with topic embedding features will be helpful in predicting persuasion since similar strategies may be effective for related topics. We further define topic semantics features to encode how focused vs. divergent each of the arguments made by the debaters is given the dis- cussion topic, similar to …

Content truncated to 500 words