A Knowledge-Enriched Model for Emotional Conversation Generation

In this poster, we propose a knowledge-enriched emotional conversation generation model (KE-EGM) that can ensure high quality content and focus on the impact of emotional factors during the conversation. First, we apply a multi-embedding fusion layer to provide this model with the token-level and sentence-level understanding. Then, the emotion flow attention mechanism combines flow emotion state and attention mechanism to learn and capture emotional information during the conversation dynamically. Finally, the multi-objective optimization mechanism is introduced to detect and generate fine-grained emotional responses. The experimental results show that KE-EGM outperforms several baselines not only in the content aspect but also in the emotional aspect.