CAT: A Contextualized Conceptualization and Instantiation Framework for Commonsense Reasoning

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 13111–13140 July 9-14, 2023 ©2023 Association for Computational Linguistics CAT: A Contextualized Conceptualization and Instantiation Framework for Commonsense Reasoning Weiqi Wang1∗, Tianqing Fang1∗, Baixuan Xu1, Chun Yi Louis Bo1, Yangqiu Song1, Lei Chen1,2 1Department of Computer Science and Engineering, HKUST, Hong Kong SAR, China 2Information Hub, HKUST (GZ), Guangzhou, China {wwangbw, tfangaa}@cse.ust.hk, bxuan@connect.ust.hk {cybo, yqsong}@cse.ust.hk, leichen@ust.hk Abstract Commonsense reasoning, aiming at endowing machines with a human-like ability to make sit- uational presumptions, is extremely challeng- ing to generalize. For someone who barely knows about meditation, while is knowledge- able about singing, he can still infer that medi- tation makes people relaxed from the existing knowledge that singing makes people relaxed by first conceptualizing singing as a relaxing event and then instantiating that event to medita- tion. This process, known as conceptual induc- tion and deduction, is fundamental to common- sense reasoning while lacking both labeled data and methodologies to enhance commonsense modeling. To fill such a research gap, we pro- pose CAT (Contextualized ConceptuAlization and InsTantiation), a semi-supervised learning framework that integrates event conceptualiza- tion and instantiation to conceptualize common- sense knowledge bases at scale. Extensive ex- periments show that our framework achieves state-of-the-art performances on two concep- tualization tasks, and the acquired abstract commonsense knowledge can significantly im- prove commonsense inference modeling. Our code, data, and fine-tuned models are pub- licly available at https://github.com/HKUST- KnowComp/CAT. 1 Introduction “Concepts are the glue that holds our mental world together.”– Murphy (2004) Commonsense reasoning is a crucial ability for machines to make situational presumptions and draw inferences from the knowledge that reflects our humans’ understanding of situations and com- mon facts (Davis, 1990; Davis and Marcus, 2015). It has gained increasing popularity in the Natural Language Processing (NLP) community with the emergence of CommonSense Knowledge Bases (CSKB) (Sap et al., 2019a; Speer et al., 2017; ∗ Equal Contribution PersonX watches football game, as a result, PersonX will: feel relaxed PersonX plays with his dog, as a result, PersonX will: be happy and relaxed PersonX [observe] as a result, PersonX will: feel relaxed PersonX [relaxing event] as a result, PersonX will: feel relaxed Conceptualization (watches football game → relaxing event) Instantiation (relaxing event → plays with his dog) Abstract Knowledge Aided Reasoning Conceptualization (watches football game → observe) Wrong Abstract Knowledge Figure 1: A demonstration of commonsense reasoning on an unknown situation, PersonX plays with his dog, with the aid of abstract commonsense knowledge. De- contextualized conceptualization, such as observe, may yield wrong abstract commonsense knowledge that can- not be instantiated within the corresponding context. Hwang et al., 2021) and large language mod- els (Bosselut et al., 2019; Rajani et al., 2019; Liu et al., 2022b; Su et al., 2022; Yu et al., 2022b). However, when encountering situations beyond the data given, more abstract background knowledge must be acquired and generalized to assist the rea- soning (Tenenbaum et al., 2011), and language models trained with an …

Content truncated to 500 words