Emoji-Inserted Emotional Dialogue Generation with Emoji-Enriched Knowledge Graph

The quality of automatic dialogue systems is getting higher and higher. Most existing works express emotions through plain texts only, which are not enough to represent the true emotion. Emojis, a cenozonic expression, have been used widely on the Internet. There is a huge amount of value hidden in emojis, not only as external knowledge, but also making the expression of emotions more intuitive. In this paper, we endeavor to design an automatic Emojis-inserted open domain emotion controllable dialogue generation system with Emoji-Enriched Knowledge Graph (Emo-EEKG). The model enriches and extends the responses from both textual and non-textual perspectives. From the textual perspective, we construct an Emoji-Enriched Common-sense Knowledge Graph (EE-CKG). Then, the model is encouraged to generate more emotional words by embedding the emotional context knowledge subgraphs. From the non-textual perspective, we design the emoji generator to select emojis and the context judger to classify the sentences. The effectiveness of the proposed approach is verified by experimentally comparing with traditional models on a Twitter dataset. Our model can generate emotional responses with automatic emoji insertion and common-sense.