Empathetic response generation via commonsense knowledge and emotional intensity

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Empathetic response generation aims to generate an appropriate response by comprehending the speaker’s emotional feelings. Previous approaches mainly focus on utilizing the speaker’s emotion for generating empathetic responses. However, these approaches usually treat emotion as a static variable and overlook the fact that the impact of emotion on the words in response is dynamic, which could lead to inappropriate responses. To address this issue, we propose a fusion of Commonsense Knowledge and Emotional Intensity (CKEI) method for Empathetic Response Generation. We first augment the dialogue context with commonsense knowledge during the encoding process, which is crucial for improving the accuracy of emotion label prediction and the relevance of the response to the dialogue context. Then, we use a two-stage decoding process to generate responses, which is more sensitive to the dynamic impact of emotion on the words in response. In the first decoding process, we predict the emotional intensity of each word in the response. After combining the emotional intensity with the embedding of the predicted emotion label, we obtain the emotional representation for each word in the response, which also represents how emotion impacts the words. Finally, we pass the emotional representation to the second decoding process to guide the generation of the response. Extensive experimental results on the Empathetic Dialogues dataset demonstrate that CKEI outperforms the baselines in both automatic and human evaluations and can generate more relevant and empathetic responses.