A Simple Contrastive Learning Framework for Interactive Argument Pair Identification via Argument-Context Extraction

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10027 - 10039 December 7-11, 2022 ©2022 Association for Computational Linguistics A Simple Contrastive Learning Framework for Interactive Argument Pair Identification via Argument-Context Extraction Lida Shi1, Fausto Giunchiglia1,2,3, Rui Song1, Daqian Shi3, Tongtong Liu2, Xiaolei Diao3, Hao Xu1,2,∗ 1School of Artificial Intelligence, Jilin University 2College of Computer Science and Technology, Jilin University 3DISI, University of Trento {shild21,songrui20,liutt20}@mails.jlu.edu.cn, xuhao@jlu.edu.cn {fausto.giunchiglia,daqian.shi,xiaolei.diao}@unitn.it Abstract Interactive argument pair identification is an emerging research task for argument mining, aiming to identify whether two arguments are interactively related. It is pointed out that the context of the argument is essential to im- prove identification performance. However, current context-based methods achieve limited improvements since the entire context typically contains much irrelevant information. In this paper, we propose a simple contrastive learn- ing framework to solve this problem by ex- tracting valuable information from the context. This framework can construct hard argument- context samples and obtain a robust and uni- form representation by introducing contrastive learning. We also propose an argument-context extraction module to enhance information ex- traction by discarding irrelevant blocks. The experimental results show that our method achieves the state-of-the-art performance on the benchmark dataset. Further analysis demon- strates the effectiveness of our proposed mod- ules and visually displays more compact se- mantic representations. The code is available at GitHub 1. 1 Introduction Computational argumentation, as a branch of natu- ral language understanding, has become a new re- search field. Existing work can be divided into two categories (Asterhan and Schwarz, 2007): mono- logical argumentation and dialogical argumenta- tion. Monological argumentation is the scenario for one participant, such as RCT (Mayer et al., 2020), student essays (Stab and Gurevych, 2014) and user comments (Niculae et al., 2017). The researchers focus on topics like argumentation (ar- gument) mining (Galassi et al., 2018; Morio et al., ∗* Corresponding author 1https://github.com/shilida/CL_Interactive_ Argument_Pair_Identification Figure 1: An instance in the dataset. Each instance in- cludes six arguments: a quotation and its corresponding five candidate replies. Additionally, the task provides contextual information for each argument. For this task, the model needs to identify whether the quotation and the reply are interactively related. Only one of the five candidate replies is correct. The arguments are repre- sented in green font, and the context is expressed in black. 2020; Jo et al., 2019; Ruiz-Dolz et al., 2021), argu- ment assessment (Anne et al., 2020; Skitalinskaya et al., 2021), and argument reasoning (Botschen et al., 2018; Habernal et al., 2018; Ruiz-Dolz et al., 2021) for this sort of study. Recently, researchers have been paying great attention to dialogical ar- gumentation, since online forums have become the primary medium for argumentation and discussion. People can express themselves on the network 10027 Figure 2: Four heatmaps of the semantic similarity be- tween the argument and each sentence in its context. The horizontal coordinate is the number of sentences, and the vertical coordinate is the range of semantic sim- ilarity. (a) quotation-context (b) positive relpy-context (c) …

Content truncated to 500 words