Neural Response Generation with Relevant Emotions for Short Text Conversation

Human conversations are often embedded with emotions. To simulate human conversations, the response generated by a chat-bot not only has to be topically relevant to the post, but should also carry an appropriate emotion. In this paper, we conduct analysis based on social media data to investigate how emotions influence conversation generation. Based on observation, we propose methods to determine the appropriate emotions to be included in a response and to generate responses with the emotions. The encoder-decoder architecture is extended to incorporate emotions. We propose two implementations which train the two steps separately or jointly. An empirical study on a public dataset from STC at NTCIR-12 shows that our models outperform both a retrieval-based method and a generation model without emotion, indicating the importance of emotions in short text conversation generation and the effectiveness of our approach.