ArgLegalSumm: Improving Abstractive Summarization of Legal Documents with Argument Mining

Proceedings of the 29th International Conference on Computational Linguistics, pages 6187–6194 October 12–17, 2022. 6187 ArgLegalSumm: Improving Abstractive Summarization of Legal Documents with Argument Mining Mohamed Elaraby, Diane Litman University of Pittsburgh Pittsburgh, PA, USA {mse30,dlitman}@pitt.edu Abstract A challenging task when generating summaries of legal documents is the ability to address their argumentative nature. We introduce a simple technique to capture the argumentative struc- ture of legal documents by integrating argu- ment role labeling into the summarization pro- cess. Experiments with pretrained language models show that our proposed approach im- proves performance over strong baselines. 1 Introduction Abstractive summarization has made great progress by leveraging large pretrained language models such as BART (Lewis et al., 2020), T5 (Raffel et al., 2020), Pegasus (Zhang et al., 2020), and Long- former (Beltagy et al., 2020). These models lever- age large scale datasets such as CNN-DailyMail (Hermann et al., 2015), PubMed (Cohan et al., 2018), and New York Times (Sandhaus, 2008). Unlike news and scientific texts, which contain specific formatting such as topic sentences and ab- stracts, legal cases contain implicit argument struc- ture spreading across long texts (Xu et al., 2021). Current abstractive summarization models do not take into account the argumentative structure of the text, which poses a challenge towards effective abstractive summarization of legal documents. In this work, we bridge the gap between prior research focusing on summarizing legal documents through extracting argument roles of legal text (Grover et al., 2003; Xu et al., 2021; Saravanan and Ravindran, 2010), and prior research focused on producing abstractive summaries of legal text (Feijo and Moreira, 2019; Bajaj et al., 2021). Our work proposes a technique that blends argument role mining and abstractive summarization, which hasn’t been explored extensively in the literature. Figure 1 describes the main flow of our approach, which decomposes the summarization process into two tasks. First, each sentence in the document is Figure 1: Overview of our approach. assigned an argument role by using an independent model. Then, the predicted roles are blended with the original document’s sentences and fed into a se- quence to sequence-based abstractive summarizer. Our contributions are as follows: (a) We propose a simple technique to create an argument-aware neural abstractive summarizer. (b) We show the effectiveness of this technique in improving legal document summarization. (c) We make our code 1 and argument role annotations freely available2. 2 Related Work Legal Document Summarization. Prior research has mainly focused on extractive techniques (Gal- gani et al., 2015; Anand and Wagh, 2019; Jain et al., 2021), exploiting features such as the document structure and prior knowledge of legal terms to ex- tract salient sentences that represent the summary of the legal text. Recent research has also shifted gears to abstractive techniques due to their superi- ority to extractive methods on automatic measures such as ROUGE (Feijo and Moreira, 2019). These abstractive techniques benefited from neural mod- els such as pointer generator networks (See et al., 2017) (utilized in legal public opinion summariza- tion (Huang et al., 2020)) …

Content truncated to 500 words