Conclusion-based Counter-Argument Generation

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 957–967 May 2-6, 2023 ©2023 Association for Computational Linguistics Conclusion-based Counter-Argument Generation Milad Alshomary Institute of Artificial Intelligence Leibniz University Hannover m.alshomary@ai.uni-hannover.de Henning Wachsmuth Institute of Artificial Intelligence Leibniz University Hannover h.wachsmuth@ai.uni-hannover.de Abstract In real-world debates, the most common way to counter an argument is to reason against its main point, that is, its conclusion. Existing work on the automatic generation of natural language counter-arguments does not address the relation to the conclusion, possibly because many arguments leave their conclusion implicit. In this paper, we hypothesize that the key to effective counter-argument generation is to ex- plicitly model the argument’s conclusion and to enforce that the stance of the generated counter is opposite to that conclusion. In particular, we propose a multitask approach that jointly learns to generate both the conclusion and the counter of an input argument. The approach employs a stance-based ranking component that selects the counter from a diverse set of generated can- didates whose stance best opposes the gener- ated conclusion. In both automatic and man- ual evaluation, we provide evidence that our approach generates more relevant and stance- adhering counters than strong baselines. 1 Introduction Given an argument, a valid counter-argument to it should be relevant to the topic discussed by the argument while opposing to its conclusion’s stance. Countering the opponent’s arguments in a debate effectively is key to winning the debate (Zhang et al., 2016). While some counter-arguments attack an argument’s premises or their connection to the conclusion, the most common attack is to directly rebut the argument’s conclusion (Walton, 2009). The automatic countering of natural language arguments is one of the most challenging tasks in the area of computational argumentation. Prior research has addressed the task through retrieval (Wachsmuth et al., 2018; Orbach et al., 2020) or generation-based approaches (Hua and Wang, 2018; Hidey and McKeown, 2019). By concept, the former requires the presence of suitable counter- arguments in a predefined collection, limiting its Conclusion (title): Purchasing meat encourages animal abuse. Premises (post): All meat, to my knowledge, is obtained by raising animals in cramped quarters and slaughtering them as soon as they are fully grown. The only exception i can think of is perhaps when you go into the woods and hunt food for yourself in which case the animal has lived an undisturbed life and is put down by humane means compared to how it happens in nature. However, this is, of course, time intensive, requires skill, expensive, and thus is of course not how the vast majority of meat is obtained. Table 1: An example argument (conclusion + premises) taken from Reddit ChangeMyView, showing how the conclusion is mentioned implicitly only in the body. flexibility. Existing generation-based approaches, on the other hand, either consider a single claim as input or do not model the relation between premise and conclusion in the input argument. In previous work, we have studied the task of counter-argument …

Content truncated to 500 words