Argument Pair Extraction via Attention-guided Multi-Layer Multi-Cross Encoding

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 6341–6353 August 1–6, 2021. ©2021 Association for Computational Linguistics 6341 Argument Pair Extraction via Attention-guided Multi-Layer Multi-Cross Encoding Liying Cheng∗1,2 Tianyu Wu1 Lidong Bing2 Luo Si2 1Singapore University of Technology and Design 2DAMO Academy, Alibaba Group {liying.cheng, l.bing, luo.si}@alibaba-inc.com tianyu wu@alumni.sutd.edu.sg Abstract Argument pair extraction (APE) is a research task for extracting arguments from two pas- sages and identifying potential argument pairs. Prior research work treats this task as a se- quence labeling problem and a binary clas- sification problem on two passages that are directly concatenated together, which has a limitation of not fully utilizing the unique characteristics and inherent relations of two different passages. This paper proposes a novel attention-guided multi-layer multi-cross encoding scheme to address the challenges. The new model processes two passages with two individual sequence encoders and updates their representations using each other’s repre- sentations through attention. In addition, the pair prediction part is formulated as a table- filling problem by updating the representations of two sequences’ Cartesian product. Further- more, an auxiliary attention loss is introduced to guide each argument to align to its paired ar- gument. An extensive set of experiments show that the new model significantly improves the APE performance over several alternatives 1. 1 Introduction Mining argumentation structures within a corpus is a crucial task in argument mining research field (Palau and Moens, 2009). There are usually two main components in learning natural language argu- ment structures: (1) detecting argumentative units, (2) predicting relations between the identified argu- ments. It has been widely studied by natural lan- guage processing (NLP) researchers (Cabrio and Villata, 2018) and applied to domains such as: web debating platforms (Boltuˇzi´c and ˇSnajder, 2015; Swanson et al., 2015; Chakrabarty et al., 2019), ∗Liying Cheng is under the Joint Ph.D. Program between Alibaba and Singapore University of Technology and Design. 1Our code and data are available at https://github. com/TianyuTerry/MLMC. persuasive essays (Stab and Gurevych, 2014; Pers- ing and Ng, 2016), social media (Abbott et al., 2016), etc. Unlike traditional argument extraction tasks that are mainly from monologues, Cheng et al. (2020) propose a new task - argument pair extrac- tion (APE) from two passages in a new domain, namely peer review process, focusing on exploit- ing the interactions between reviewer comments and author rebuttals. As shown in Figure 1, APE task aims to extract the argument pairs from two passages. Specific suggestions, questions or chal- lenges in reviews are considered as review argu- ments. Response sentences that answer or explain the specific review argument are its paired rebut- tal arguments. For example in the pink area, the reviewer points out the lack of literature review in submission (i.e., review sentences 11-12). As a response, the authors argue that they select the literature based on the special focus of their work (i.e., rebuttal sentence 6-7). Similar to the two components in the traditional argumentation structure mining, …

Content truncated to 500 words