(ICCC 2021) Conversational Humor Identification based on Adversarial Learning on Chinese Sitcoms

Abstract: Humor computing aims to recognize, interpret, and generate humorous expressions based on computational models, which has become one of the key issues in the field of natural language processing. Humor computing is essential for artificial intelligence to make machines more humane. However, most research focuses on one-liner, and few research has paid attention to humor identification in conversations, that is conversational humor identification. Conversational humor identification faces the following challenges: 1.dataset is difficult to be constructed; 2.key features indicating conversational humorous expressions are hard to be captured by traditional ways. Our work is based on a Chinese sitcom dataset, which consists of all the dialogues of the sitcom. Punchlines referring to the part of dialogues play a role in making people laugh are also annotated in the dataset. Conversational humor identification is to identify the punchline. To do this, we propose a humor identification model based on Adversarial Learning, where the generator is able to produce sequences similar to punchlines, while the discriminator can learn the features to distinguish punchlines from non-humor dialogue. In this way, when given an input dialogue, the discriminator will classify whether it is a punchline for each sentence. We conduct experiments to assess the performance of our model, and we also compare the experimental results with some strong baselines. There are 25.4% improvement over BERT with respect of F1-score. Moreover, we also design ablation experiment to analyze different factors in conversational humor identification. And we discover that the generator plays a more important role in the model.

Authors: Wenbo Shang, Jing Wei, Runhui Song, Yan Xu and Binyang Li (University of International Relations, China)

Email: nan_zhi713@163.com, sunhuisong@163.com, xuyan_0158@163.com, beyondlee1982@gmail.com