Informative and diverse emotional conversation generation with variational recurrent pointer-generator

1 Introduction and main contributions The controllable multi-turn emotional response generation task aims to understand the feelings and intentions according to the conversation history between two speakers,and then generate emotionally consistent and empathic responses based on the specific emotion category for smooth human-machine communication[1].The traditional methods[1–3]on this task only focused on enriching the response with desired emotion,which,however,can only generate safe and dull emotional responses.Inspired by the psychology theory on the emotional trigger expressions[4],in this paper,we conjecture that taking the expressions into consideration can refine conversation generation and enhance user satisfaction.Specifically,the emotional trigger expressions can be defined as:the entity-related words and phrases which can provoke and cause the desired emotion of any response.