Mining Effective Features Using Quantum Entropy for Humor Recognition

Findings of the Association for Computational Linguistics: EACL 2023, pages 2048–2053 May 2-6, 2023 ©2023 Association for Computational Linguistics Mining Effective Features Using Quantum Entropy for Humor Recognition Yang Liu and Yuexian Hou College of Intelligence and Computing, Tianjin University, Tianjin, China {lauyon,yxhou}@tju.edu.cn Abstract Humor recognition has been studied with sev- eral different methods in the past years. How- ever, existing studies on humor recognition do not understand the mechanisms that gener- ate humor. In this paper, inspired by the in- congruity theory, any joke can be divided in- to two components (the setup and the punch- line). Both components have multiple pos- sible semantics, and there is an incongruous relationship between them. We use density matrices to represent the semantic uncertain- ty of the setup and the punchline, respective- ly, and design Quantum Entropy Uncertain- ty (QE-Uncertainty) and Quantum Entropy In- congruity (QE-Incongruity) with the help of quantum entropy as features for humor recog- nition. The experimental results on the Se- mEval2021 Task 7 dataset show that the pro- posed features are more effective than the baselines for recognizing humorous and non- humorous texts. 1 Introduction Humor is one of the most distinctive features of human behavior and a sign of mental maturity (Pasquali, 1990). The study of humor has re- ceived extensive attention in the fields of linguistics, philosophy, psychology, and sociology (Mihalcea et al., 2010). Computational humor is of particular interest, with the potential to transform computers into creative and motivational tools (Nijholt et al., 2003). This paper restricts research to humor recogni- tion in computational humor, which aims to rec- ognize whether a piece of text is humorous. As shown in Figure 1, a joke usually includes two components: the setup and the punchline. The reader generates an expectation of the following text (the punchline) based on the content of the setup, and if the following text violates the reader’s expectation, humor is generated, and vice versa. ������������������������������������� ������������������������������� ������������������������������� ��������������������������� ������������������������������������������������ ���������������������������� �������������������������� ������������������� ������������������������� �������������������� ����� ��������� Figure 1: A humor and non-humor example containing the setup and the punchline. In fact, the incongruity theory of humor can ex- plain the above process of producing humor. The incongruity theory states that humor is generated because a thing (the setup) has multiple underlying concepts, and there is an incongruity between the concept involved in the situation and the real object it represents (the punchline). Features based on semantic similarity (Mihalcea et al., 2010; Yang et al., 2015) and word associa- tion (Liu et al., 2018; Cattle and Ma, 2018) have achieved certain results, but they lack considera- tion of humorous mechanisms. Xie et al. (2021) calculated the uncertainty and the surprisal values of the joke with the help of the GPT-2. But they did not model the semantic incongruity between the setup and the punchline. While the above ap- proaches are somewhat effective, the incongruity theory requires us to model semantic uncertain- ty and the incongruity between the setup and the punchline. We take inspiration …

Content truncated to 500 words