This paper introduces TalkFunny, a large-scale Chinese explainable humor response dataset with chain-of-humor and humor mind map annotations. The dataset consists of 4,116 high-quality context-response pairs crawled from various platforms. Each pair is annotated with a chain-of-humor field that explains how the humorous text is generated and a humor mind map field that reveals the underlying knowledge and logic for generating the humor response. The dataset also includes auxiliary tasks such as humor sentiment-style classification and humor rewriting to further enhance the humorous response ability of pre-trained language models (PLMs). The paper evaluates the effectiveness of the dataset and auxiliary tasks in improving the humor response performance of PLMs. The results show that the proposed dataset and auxiliary tasks significantly improve the ability of PLMs to generate humorous responses. The paper also discusses the challenges in humor generation, including the lack of humor corpus and the need for rich knowledge and commonsense. The dataset and auxiliary tasks are designed to help PLMs better understand and generate humor. The paper also presents a case study and error analysis of the generated responses. The results demonstrate that the proposed methods can effectively help PLMs generate humorous responses. The paper concludes that humor response generation is a challenging task in NLP and that the TalkFunny dataset provides a valuable resource for future research in this area.This paper introduces TalkFunny, a large-scale Chinese explainable humor response dataset with chain-of-humor and humor mind map annotations. The dataset consists of 4,116 high-quality context-response pairs crawled from various platforms. Each pair is annotated with a chain-of-humor field that explains how the humorous text is generated and a humor mind map field that reveals the underlying knowledge and logic for generating the humor response. The dataset also includes auxiliary tasks such as humor sentiment-style classification and humor rewriting to further enhance the humorous response ability of pre-trained language models (PLMs). The paper evaluates the effectiveness of the dataset and auxiliary tasks in improving the humor response performance of PLMs. The results show that the proposed dataset and auxiliary tasks significantly improve the ability of PLMs to generate humorous responses. The paper also discusses the challenges in humor generation, including the lack of humor corpus and the need for rich knowledge and commonsense. The dataset and auxiliary tasks are designed to help PLMs better understand and generate humor. The paper also presents a case study and error analysis of the generated responses. The results demonstrate that the proposed methods can effectively help PLMs generate humorous responses. The paper concludes that humor response generation is a challenging task in NLP and that the TalkFunny dataset provides a valuable resource for future research in this area.