Contrastive Argument Summarization Using Supervised and Unsupervised Machine Learning

Controversial topics like abortion or capital punishment inherently lack of correct answers or the right way to deal with. Thus, in order to find what is true, what is good, or what should be done, the involved parties need to debate. For the purpose of forming an opinion on a controversial topic someone needs to take in a lot of arguments on that topic to gather information which can be a time-consuming process. To increase efficiency, someone can use an argument search engine to quicken the retrieval of relevant arguments. Although the usage of such a service reduces the time to find arguments, there is still a lot of textual data that needs to be read. To this end, computational summarization approaches for arguments can limit the necessary time for information review by generating short snippets capturing the main gist of each argument. Yet, we suggest that approaches that consider one argument at a time show potential for further improvement in terms of efficiency during information review. In fact, arguments on the same topic, like those retrieved by a search engine for a certain query, partially cover the same content, e. g. arguments regarding the death penalty probably use deterrence as a point in favor of it. However, if the same aspect is central in multiple arguments, their snippets reflect this, which leads to redundancy among the snippets. Consequently, someone interested in gathering information on a controversial topic does not necessarily find new information in each snippet he or she reads. We introduce the task of Contrastive Argument Summarization (CAS) which addresses the aforementioned problem regarding existing argument summarization. An approach that addresses CAS aims to produce contrastive snippets for each argument in a set of topic-related arguments. A contrastive snippet should represent the main gist of its argument, it should account for the argumentative nature of the text, and it should be dissimilar to the other topic-related arguments in order to reduce redundancy among the snippets. We propose two approaches addressing CAS, namely an extended version of the LexRank derivation by Alshomary et al. (2020), and an advancement of the work by Bista et al. (2020). Additionally, we develop two automatic measures to assess to which extent the snippets of one set are opposed. For evaluation, we compile a corpus using the args.me search engine Wachsmuth et al. (2017b) to come close to the suggested area of application. Moreover, we conduct a manual annotation study to assess approaches’ effectiveness. We find that the graph-based approach is superior when it comes to contrastiveness (i. e. snippets being dissimilar to topic-related arguments), and that the second approach outperforms the previous one and the unmodified version of Alshomary et al. (2020) when it comes to representativeness (i. e. snippets capturing the main gist of an argument).