Employing Argumentation Knowledge Graphs for Neural Argument Generation

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4744–4754 August 1–6, 2021. ©2021 Association for Computational Linguistics 4744 Employing Argumentation Knowledge Graphs for Neural Argument Generation Khalid Al-Khatib1 Lukas Trautner2 Henning Wachsmuth3 Yufang Hou4 Benno Stein5 1 Leipzig University, Germany, khalid.alkhatib@uni-leipzig.de 2 University of Erlangen-Nuremberg, Germany, lukas.trautner@fau.de 3 Paderborn University, Germany, henningw@upb.de 4 IBM Research Europe, Ireland, yhou@ie.ibm.com 5 Bauhaus-Universität Weimar, Germany, benno.stein@uni-weimar.de Abstract Generating high-quality arguments, while be- ing challenging, may benefit a wide range of downstream applications, such as writing as- sistants and argument search engines. Moti- vated by the effectiveness of utilizing knowl- edge graphs for supporting general text gener- ation tasks, this paper investigates the usage of argumentation-related knowledge graphs to control the generation of arguments. In partic- ular, we construct and populate three knowl- edge graphs, employing several compositions of them to encode various knowledge into texts of debate portals and relevant paragraphs from Wikipedia. Then, the texts with the encoded knowledge are used to fine-tune a pre-trained text generation model, GPT-2. We evaluate the newly created arguments manually and auto- matically, based on several dimensions impor- tant in argumentative contexts, including ar- gumentativeness and plausibility. The results demonstrate the positive impact of encoding the graphs’ knowledge into debate portal texts for generating arguments with superior quality than those generated without knowledge. 1 Introduction Arguments are our means to build stances on con- troversial topics, to persuade others, or to negotiate. Automatic argument generation has the potential to effectively support such tasks: it may not only regenerate known arguments but also uncover new facets of a topic. Existing argument generation ap- proaches work either in an end-to-end fashion (Hua and Wang, 2018) or they are controlled with respect to the argument’s topic, aspects, or stance (Gretz et al., 2020; Schiller et al., 2021). In contrast, no approach integrates external knowledge into the generation process so far, even though knowledge graphs have been shown to be useful for support- ing text generation models in other areas (Koncel- Kedziorski et al., 2019a; Ribeiro et al., 2020). Previous research has proposed argumentation knowledge graphs (AKGs) that model supporting and attacking interactions between concepts (Al- Khatib et al., 2020). Such an AKG may assist argument generation models in different ways. For example, meaningful prompts on controversial top- ics can be constructed from an AKG with sim- ple hand-defined rules, such as ‘geoengineering re- duces atmospheric greenhouse gas’ for generating an argument on ‘geoengineering.’ Alternatively, an AKG may be employed to control the generation, making arguments adhere to knowledge covered in the graph. We hypothesize this to be particularly beneficial for the quality of arguments in terms of factuality, the richness of evidence, and similar. This paper concentrates on such controlled ar- gument generation, investigating for the first time the ability to generate high-quality and content-rich arguments by integrating knowledge from AKGs into standard neural-based generation models. To this end, we exploit multiple manually and …

Content truncated to 500 words