Generating Responses with Emotional Scoring

Emotional expression is regarded as an important aspect of human intelligence, which inspires a lot of researchers working on enabling chatbots to give responses with controlled emotion. Deep learning method is widely applied to this task to generate responses according to the input post and emotion label. In previous works, the sentiment dictionary is widely used in gen-erating emotional responses. However, most external dictionaries only provide sentiment polarities, which are coarse-grained. In this article, we propose an emotional response generation model based on an emotion scoring module and a fine-grained sentiment dictionary. The emotion scoring module is pre-trained to give sentimental scores according to the target fine-grained emotion label, which is used to guide response generation. We introduced dynamic emotional embeddings instead of fixed ones used in previous works. As for the sentiment dictionary, we use a fine-grained sentiment dictionary of different emotion categories. At the same time, to build the one-to-one correlation between post and response, we acquire conditional variational auto-encoder to acquire a latent variable of responses. Experiments demonstrate that our model achieves state-of-the-art performance in generating responses with specified emotions.