Improving Commonsense Contingent Reasoning by Pseudo-data and Its Application to the Related Tasks

Proceedings of the 29th International Conference on Computational Linguistics, pages 812–823 October 12–17, 2022. 812 Improving Commonsense Contingent Reasoning by Pseudo-data and its Application to the Related Tasks Kazumasa Omura Sadao Kurohashi Graduate School of Informatics, Kyoto University Yoshida-honmachi, Sakyo-ku, Kyoto, 606-8501, Japan {omura, kuro}@nlp.ist.i.kyoto-u.ac.jp Abstract Contingent reasoning is one of the essential abilities in natural language understanding, and many language resources annotated with con- tingent relations have been constructed. How- ever, despite the recent advances in deep learn- ing, the task of contingent reasoning is still difficult for computers. In this study, we focus on the reasoning of contingent relation between basic events. Based on the existing data con- struction method, we automatically generate large-scale pseudo-problems and incorporate the generated data into training. We also in- vestigate the generality of contingent knowl- edge through quantitative evaluation by per- forming transfer learning on the related tasks: discourse relation analysis, the Japanese Wino- grad Schema Challenge, and the JCommon- senseQA. The experimental results show the effectiveness of utilizing pseudo-problems for both the commonsense contingent reasoning task and the related tasks, which suggests the importance of contingent reasoning. 1 Introduction Contingency is the relation between two events, one being an action or state and the other being likely to happen after it. We humans reason contin- gent relation between events on a daily basis. For instance, when we read text, we unconsciously in- fer what happens next to deepen our understanding. In conversations, we guess the next topic from the utterance of the opponent to make a contextual and natural response. Thus, the ability to reason con- tingent relation between events is essential when it comes to natural language understanding (NLU). Recently, many studies have built language re- sources for contingent reasoning (Roemmele et al., 2011; Mostafazadeh et al., 2016; Zellers et al., 2018; Sap et al., 2019a; Hwang et al., 2021). These resources focus on basic events and evaluate some kind of commonsense reasoning ability. Although the fundamental linguistic capabilities of comput- I’m hungry, so a. I’m gonna be absent from school. b. I refrain from strenuous exercise. c. I have a meal at a family restaurant. d. I leave home. Figure 1: Example from KUCI (English translated ver- sion). KUCI is a Japanese QA dataset containing 104k multiple-choice questions regarding contingent relation between basic events. The correct choice is bolded. ers, such as question answering, have greatly im- proved with progress in deep learning, several stud- ies have empirically demonstrated they still have difficulty in commonsense reasoning (Talmor et al., 2019; Sap et al., 2019b; Talmor et al., 2021). In this study, we aim at two objectives: to im- prove commonsense contingent reasoning and to investigate the effects of learning contingent knowl- edge on the related tasks to validate the importance of contingent reasoning. To these ends, we use the Kyoto University Commonsense Inference dataset (KUCI)1. KUCI is a Japanese QA dataset with 104k multiple-choice questions that ask contingent rela- tion between basic events directly (Omura et al., 2020). An …

Content truncated to 500 words