Target Inference in Argument Conclusion Generation

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4334–4345 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 4334 Target Inference in Argument Conclusion Generation Milad Alshomary Paderborn University Department of Computer Science milad.alshomary@mail.upb.de Shahbaz Syed Leipzig University Department of Computer Science shahbaz.syed@uni-leipzig.de Martin Potthast Leipzig University Department of Computer Science martin.potthast@uni-leipzig.de Henning Wachsmuth Paderborn University Department of Computer Science henningw@upb.de Abstract In argumentation, people state premises to rea- son towards a conclusion. The conclusion con- veys a stance towards some target, such as a concept or statement. Often, the conclusion re- mains implicit, though, since it is self-evident in a discussion or left out for rhetorical reasons. However, the conclusion is key to understand- ing an argument, and hence, to any application that processes argumentation. We thus study the question to what extent an argument’s con- clusion can be reconstructed from its premises. In particular, we argue here that a decisive step is to infer a conclusion’s target, and we hypoth- esize that this target is related to the premises’ targets. We develop two complementary tar- get inference approaches: one ranks premise targets and selects the top-ranked target as the conclusion target, the other finds a new conclu- sion target in a learned embedding space using a triplet neural network. Our evaluation on cor- pora from two domains indicates that a hybrid of both approaches is best, outperforming sev- eral strong baselines. According to human an- notators, we infer a reasonably adequate con- clusion target in 89% of the cases. 1 Introduction The conclusion (or claim) of a natural language ar- gument conveys a pro or con stance towards some target, such as a controversial concept or statement (Bar-Haim et al., 2017). It is inferred from a set of premises. Conclusions are key to understanding arguments, and hence, critical for any downstream application that processes argumentation. The task of identifying conclusions has been studied inten- sively in the context of argument mining (Stab and Gurevych, 2014) and automatic essay assessment (Falakmasir et al., 2014). In genres other than es- says, however, conclusions often remain implicit, since they are clear from the context of a discussion (Habernal and Gurevych, 2015) or hidden on pur- Raising the school leaving age promotes equal opportunities. Parents who left school at a young age are also more likely to have children who leave school early. Making sure that everyone gets the same amount of time at school promotes equality. Forcing all children to stay in school longer will help break this cycle of disadvantage. Parents who left school at a young age Making sure [...] same amount of time at school. Forcing all children to stay in school longer con pro pro Premise target identification (1) Conclusion target inference (3) Conclusion text generation Raising the school leaving age pro Stance classification (2) Stance inference tackled in this paper Conclusion Premises Figure 1: Illustration of our full model of generating an argument’s conclusion from its premises. This paper focuses …

Content truncated to 500 words