A Structure-Aware Argument Encoder for Literature Discourse Analysis

Proceedings of the 29th International Conference on Computational Linguistics, pages 7093–7098 October 12–17, 2022. 7093 A Structure-Aware Argument Encoder for Literature Discourse Analysis Yinzi Li1,†, Wei Chen1,†, Zhongyu Wei1,2,∗, Yujun Huang1, Chujun Wang3 Siyuan Wang1, Qi Zhang3, Xuanjing Huang3 and Libo Wu1 1School of Data Science, Fudan University, China 2Research Institute of Intelligent Complex Systems, Fudan University 3Department of Environmental Science and Engineering, Fudan University, China 3School of Computer Science, Fudan University, China {yinzili20,chenwei18,zywei,yjhuang19,chujunwang20}@fudan.edu.cn; {wangsy18,qz,xjhuang,wulibo}@fudan.edu.cn; Abstract Existing research for argument representation learning mainly treats tokens in the sentence equally and ignores the implied structure in- formation of argumentative context. In this paper, we propose to separate tokens into two groups, namely framing tokens and topic ones, to capture structural information of ar- guments. In addition, we consider high-level structure by incorporating paragraph-level po- sition information. A novel structure-aware ar- gument encoder is proposed for literature dis- course analysis. Experimental results on both a self-constructed corpus and a public cor- pus show the effectiveness of our model. Re- sources are available at https://github. com/lemuria-wchen/SAE. 1 Introduction With the growing amount of scientific literature, researchers pay increasing attention to developing computational methods for analyzing scientific lit- erature (Kirschner et al., 2015; Stab and Gurevych, 2017; Green, 2018; Lauscher et al., 2018; Ac- cuosto and Saggion, 2019), aiming to identify various components of arguments automatically (Abend et al., 2009; Judea and Strube, 2017; Lukin et al., 2017; Durmus and Cardie, 2018; Lugini and Litman, 2020). Existing research focuses on constructing annotated corpus and learning rep- resentation of sentences for literature discourse analysis. They tend to treat tokens in a sentence equally and ignore the implied structure informa- tion of argumentative context (Stab and Gurevych, 2014; Wachsmuth et al., 2016; Zhang et al., 2016; Lawrence and Reed, 2017). Figure 1 shows two annotated abstracts of sci- entific literature, in which sentences are classi- fied into four types, namely background, method, result and conclusion. We have some findings. ∗Corresponding author †Equal contributions Figure 1: Two samples of annotated abstracts. Fram- ing tokens are highlighted in blue font and blue dotted lines and the rest are topic tokens. The division rule of tokens can be referred in section 3. First, tokens in sentences can be divided into two groups, i.e., topic words and framing words. Topic words provide the fundamental knowledge of this argument while framing words organize the ex- pression. Second, the same argument components often use similar framing structure in discourses across topics. For example, structures like ‘... is employed / investigated to ...’ usually appear in the method section. Third, argument components are sensitive to their positions. For example, back- ground almost always comes before method part and conclusion usually locates at the end. Moti- vated by these findings, we propose a structure- aware argument encoder (SAE) based on the trans- former to enhance the literature discourse analysis. Experimental results show the effectiveness of our proposed model both on a self-constructed corpus and a public corpus. Our contributions are two-fold: (1) we propose …

Content truncated to 500 words