Similarity-weighted Construction of Contextualized Commonsense Knowledge Graphs for Knowledge-intense Argumentation Tasks

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 6130–6158 July 9-14, 2023 ©2023 Association for Computational Linguistics Similarity-weighted Construction of Contextualized Commonsense Knowledge Graphs for Knowledge-intense Argumentation Tasks Moritz Plenz† Juri Opitz† Philipp Heinisch‡ Philipp Cimiano‡ Anette Frank† †Heidelberg University ‡Bielefeld University {plenz,opitz,frank}@cl.uni-heidelberg.de {pheinisch,cimiano}@techfak.uni-bielefeld.de Abstract Arguments often do not make explicit how a conclusion follows from its premises. To com- pensate for this lack, we enrich arguments with structured background knowledge to support knowledge-intense argumentation tasks. We present a new unsupervised method for con- structing Contextualized Commonsense Knowl- edge Graphs (CCKGs) that selects contextu- ally relevant knowledge from large knowl- edge graphs (KGs) efficiently and at high qual- ity. Our work goes beyond context-insensitive knowledge extraction heuristics by computing semantic similarity between KG triplets and textual arguments. Using these triplet simi- larities as weights, we extract contextualized knowledge paths that connect a conclusion to its premise, while maximizing similarity to the argument. We combine multiple paths into a CCKG that we optionally prune to reduce noise and raise precision. Intrinsic evaluation of the quality of our graphs shows that our method is effective for (re)constructing human explana- tion graphs. Manual evaluations in a large-scale knowledge selection setup confirm high recall and precision of implicit CSK in the CCKGs. Finally, we demonstrate the effectiveness of CCKGs in a knowledge-insensitive argument quality rating task, outperforming strong base- lines and rivaling a GPT-3 based system.1 1 Introduction Computational argumentation is a growing field with relevant applications, such as argument re- trieval (Wachsmuth et al., 2017b; Bondarenko et al., 2021), argument analysis (Feng and Hirst, 2011; Janier et al., 2014; Wachsmuth et al., 2017a; Jo et al., 2020; Opitz et al., 2021) or generation (Schiller et al., 2021; Alshomary et al., 2021; Heinisch et al., 2022a). Argumentation requires deep understanding of argumentative statements 1Our code and data are available at https://github.com/Heidelberg-NLP/CCKG Figure 1: CCKG connecting the premise "A person is unhappy if she is dissatisfied with her body." to the con- clusion "Plastic surgery raises patients’ self esteem and allows them to lead normal happy lives." Concepts asso- ciated with premise and conclusion are colored in violet and orange, respectively. The graph makes explicit that plastic surgery causes looking better, which in turn causes happiness, which is distinct from dissatisfaction. and how they relate to each other. Often, common- sense knowledge (CSK) is needed to understand how a premise connects to its conclusion, as these connections are often left implicit, as shown in Figure 1. While humans can easily infer implied knowledge, for machines extra mechanisms are needed to inject missing knowledge to better solve argumentative tasks (Moens, 2018; Becker et al., 2020; Lauscher et al., 2022; Singh et al., 2022). Methods to inject such knowledge either rely on parametric memory, where CSK is stored in the parameters of large language models (LLMs), or non-parametric memory, where CSK is stored in external knowledge bases. In the LLM approach, latent CSK can be directly exploited in …

Content truncated to 500 words