From Arguments to Key Points: Towards Automatic Argument Summarization

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4029–4039 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 4029 From Arguments to Key Points: Towards Automatic Argument Summarization Roy Bar-Haim Lilach Eden Roni Friedman Yoav Kantor Dan Lahav Noam Slonim∗ IBM Research {roybar,lilache,roni.friedman-melamed,yoavka,noams}@il.ibm.com dan.lahav@ibm.com Abstract Generating a concise summary from a large collection of arguments on a given topic is an intriguing yet understudied problem. We pro- pose to represent such summaries as a small set of talking points, termed key points, each scored according to its salience. We show, by analyzing a large dataset of crowd-contributed arguments, that a small number of key points per topic is typically sufficient for covering the vast majority of the arguments. Furthermore, we found that a domain expert can often pre- dict these key points in advance. We study the task of argument-to-key point mapping, and in- troduce a novel large-scale dataset for this task. We report empirical results for an extensive set of experiments with this dataset, showing promising performance. 1 Introduction Governments, businesses and individuals, all need to make decisions on a daily basis: “Should cannabis be legalized?”,“Should we develop this product?”, “Should I become a vegetarian?”. When making an important decision, the process typically comprises several steps: first, we gather as much information as we can about the pros and cons of the proposal under consideration. We may then summarize the collected information as a short list of the main arguments for each side. Lastly, we aim to weigh the pro and con arguments against each other to make the final decision. Where can we find relevant arguments for a given topic? In recent years, significant progress was made in the field of argument mining, auto- matic identification and extraction of argumenta- tive structures in text (Lawrence and Reed, 2020). Specifically, several works focused on topic-related argument mining from the Web or other massive corpora (Levy et al., 2017, 2018; Wachsmuth et al., ∗All authors equally contributed to this work. 2017; Stab et al., 2018a,b; Ein-Dor et al., 2020). Policy makers in governments or businesses may also conduct surveys to collect from large audi- ences arguments supporting or contesting some proposal. Each of the above methods may result in hun- dreds or thousands of arguments per topic, making it impossible for the decision maker to read and digest such large amounts of information. Several works aimed to alleviate this problem by cluster- ing together related arguments, based on different notions of relatedness, such as similarity (Reimers et al., 2019), frames (Ajjour et al., 2019), and ar- gument facets (Misra et al., 2016). These works, however, did not attempt to create a concise textual summary from the resulting clusters. In this work we propose to summarize the argu- ments supporting each side of the debate by map- ping them to a short list of talking points, termed key points. The salience of each key point can be represented by the number of its matching argu- …

Content truncated to 500 words