Offer a Different Perspective: Modeling the Belief Alignment of Arguments in Multi-party Debates

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11939 - 11950 December 7-11, 2022 ©2022 Association for Computational Linguistics Offer a Different Perspective: Modeling the Belief Alignment of Arguments in Multi-party Debates Suzanna Sia∗ Johns Hopkins University ssia1@jhu.edu Kokil Jaidka∗ National University of Singapore jaidka@nus.edu.sg Hansin Ahuja IIT Ropar hansinahuja@gmail.com Niyati Chhaya Adobe Research India nchhaya@adobe.com Kevin Duh Johns Hopkins University kevinduh@cs.jhu.edu Abstract In contexts where debate and deliberation are the norm, the participants are regularly pre- sented with new information that conflicts with their original beliefs. When required to update their beliefs (belief alignment), they may choose arguments that align with their worldview (confirmation bias). We test this and competing hypotheses in a constraint- based modeling approach to predict the win- ning arguments in multi-party interactions in the Reddit Change My View and Intelligence Squared debates datasets. We adopt a hierar- chical generative Variational Autoencoder as our model and impose structural constraints that reflect competing hypotheses about the na- ture of argumentation. Our findings suggest that in most settings, predictive models that an- ticipate winning arguments to be further from the initial argument of the opinion holder are more likely to succeed. 1 Introduction On social media, individuals are often exposed to information that conflicts with their beliefs, which may result in them experiencing cognitive disso- nance (Festinger, 1962; Festinger et al., 2017). In the context of multi-party online debates, the Com- menter (C) tries to change the Opinion Holder’s (O) point of view. We may observe a confirmation bias, when exposed to information that conflicts with their beliefs, people may seek out and favor supporting arguments that are closest to their own beliefs (Taber and Lodge, 2006; Festinger et al., 2017), leading to polarized online spaces (Bail et al., 2018). However, there is also evidence to support that novel information could work in fa- vor of changing the O’s view. Based on evidence from three different online experiments, Guess and Coppock (2020) reported that “when people are ex- posed to (new) information, they update their views in the expected or ‘correct’ direction, on average.” ∗Suzanna Sia & Kokil Jaidka co-lead this work. Which paradigm better describes the norms of online and offline debates? Work on modeling per- suasion in online forums has focused on identify- ing debate winners (Potash and Rumshisky, 2017; Zhang et al., 2016; Wang et al., 2017; Prabhakaran et al., 2013) and winning negotiation games (Keizer et al., 2017). Most of this work focuses on engi- neering or learning features to predict argument success. Previous work has suggested that to be persuasive, the commenter (C) should exert influ- ence (Hidey and McKeown, 2018), and perform effective social interaction (Wei et al., 2016; Jo et al., 2018) and language interplay (Tan et al., 2016). Many of these studies suggest that inter- play with the original opinion holder (O) would support a competing argument. However, it is not clear whether in this interaction, C should try to reflect O’s views and build …

Content truncated to 500 words