Generating emotional response by conditional variational auto-encoder in open-domain dialogue system

Abstract

An important goal in open-domain dialogue research is to make chatbot generate emotional responses given a context. To achieve this, some researchers have attempted to introduce affective information into neural dialogue models. However, these neural dialogue models containing affective information still suffer from the problem of generating safe but meaningless responses, such as I don’t know, which makes users lose interest in chatting quickly. Fortunately, the latest research has proven that conditional variational auto-encoder (CVAE) can solve this problem and enhance the responses’ diversity. In this paper, we combine affective knowledge into the CVAE-based model to generate diverse and affective responses. First, we use an affective lexicon to understand each word’s emotion in the input sentences and feed the affective vector with its embedding vector together into the CVAE-based model. Next, we construct semantic and affective loss functions, enabling the model to simultaneously learn the response’s semantic and affective distributions. Additionally, we formulate a ranking rule to help rank the candidate responses according to their syntax, semantics, and affection scores, thereby enhancing the emotion and relevance while retaining the response’s diversity. Finally, we evaluate the proposed model on the DailyDialog dataset and Reddit dataset. The experimental results show that our model can generate more emotional, diverse, and context-relevant responses compared to the baselines.