Modeling semantic and emotional relationship in multi-turn emotional conversations using multi-task learning

Recognition and expression of emotion are key factors to the success of multi-turn conversations. Emotion recognition that can help model the relationship between query and response is used to be employed in single-turn conversation models. However, little work focuses on infusing the emotional factor in multi-turn conversation generation so far. To alleviate these problems, we propose Multi-turn Emotional Conversation Model (MECM) by using multi-task learning, which improves the ability to represent emotions in multi-turn conversations. MECM is based on hierarchical latent variable model, that utilizes context hidden to sharing the common information. Besides it also contains an emotion classifier to help the model recognize the emotion in the conversation, and a conversation generator to maintain consistency of content and transformation of emotion. Experimental results show that our model significantly improves the quality of responses in terms of diversity and empathy, and keeps better performance on semantic similarity compared with baseline methods.