Exploring Discourse Structures for Argument Impact Classification

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3958–3969 August 1–6, 2021. ©2021 Association for Computational Linguistics 3958 Exploring Discourse Structures for Argument Impact Classification Xin Liu1∗ Jiefu Ou1 Yangqiu Song1 Xin Jiang2 1Department of CSE, the Hong Kong University of Science and Technology 2Huawei Noah’s Ark Lab xliucr@cse.ust.hk jouaa@connect.ust.hk yqsong@cse.ust.hk jiang.xin@huawei.com Abstract Discourse relations among arguments reveal logical structures of a debate conversation. However, no prior work has explicitly studied how the sequence of discourse relations influ- ence a claim’s impact. This paper empirically shows that the discourse relations between two arguments along the context path are essential factors for identifying the persuasive power of an argument. We further propose DISCOC to inject and fuse the sentence-level structural discourse information with contextualized fea- tures derived from large-scale language mod- els. Experimental results and extensive anal- ysis show that the attention and gate mecha- nisms that explicitly model contexts and texts can indeed help the argument impact classifica- tion task defined by Durmus et al. (2019), and discourse structures among the context path of the claim to be classified can further boost the performance. 1 Introduction It is an interesting natural language understanding problem to identify the impact and the persuasive- ness of an argument in a conversation. Previous works have shown that many factors can affect the persuasiveness prediction, ranging from textual and argumentation features (Wei et al., 2016), style fac- tors (Baff et al., 2020), to the traits of source or au- dience (Durmus and Cardie, 2018, 2019; Shmueli- Scheuer et al., 2019). Discourse relations, such as Restatement and Instantiation, among arguments reveal logical structures of a debate conversation. It is natural to consider using the discourse structure to study the argument impact. Durmus et al. (2019) initiated a new study of the influence of discourse contexts on determining ar- gument quality by constructing a new dataset Kialo. ∗ This work was done when Xin Liu was an intern at Huawei Noah’s Ark Lab. Thesis: Physical torture of prisoners is an acceptable interrogation tool. S1: Torture can help force prisoners to reveal information that could prevent attacks and save lives. O1: Torture is ineffective at getting prisoners to reveal desired information. O2: If torture is allowed, then it could easily be misused or performed in excess. S2: The knowledge that torture is acceptable and may be applied is in and of itself a strong incentive for prisoners to cooperate with their captors. S3: Interrogators and prison guards could torture prisoners solely to fulfill their own sadistic desires or out of a motivation for personal revenge. Support Support Oppose Support Oppose Result Result Contrast Result Not Impactful Impactful Impactful Restatement, Instantiation Figure 1: Example of an argument tree from Kialo. Stances, impact labels, and discourse relations are an- notated in orange, red, and violet respectively. As shown in Figure 1, it consists of arguments, im- pact labels, stances where every argument is lo- cated …

Content truncated to 500 words