The widespread development of conversational agents (chatbots) has enabled us to communicate and collaborate with different forms and functions of robots using natural language, thus facilitating a closer relationship between humans and technology. Given that chatbot services infused with domain knowledge are of great interest to not only global businesses but also academics, chatbots have in recent years become a popular research topic in the field of natural language processing. We therefore aim at improving current chatbots with the addition of natural emotions. In contrast to previous work, we intend to distinguish fine-grained emotion differences between words in order to better understand emotion expressions in sentences. Our approach infuses fine-grained emotion content into the response generation process to make the dialog more emotionally resonant. The experimental results demonstrate that this method can classify emotions more effectively. In addition, the proposed hybrid model, which consists of recurrent and convolutional neural networks with additional emotion-specific valence-arousal features, can correctly identify five emotions with a 67.89% overall F1-score. We further evaluate the subjective quality of the responses and discover that the infusion of fine-grained emotion information substantially improves the quality and fluency of automatically generated empathetic conversation. We conclude that the proposed model can greatly improve the efficiency and usability of a conversational chatbot system.
ASJC Scopus subject areas