Abstract
Towards human-like dialogue systems, current emotional dialogue approachesjointly model emotion and semantics with a unified neural network. Thisstrategy tends to generate safe responses due to the mutual restriction betweenemotion and semantics, and requires rare emotion-annotated large-scale dialoguecorpus. Inspired by the "think twice" behavior in human dialogue, we propose atwo-stage conversational agent for the generation of emotional dialogue.Firstly, a dialogue model trained without the emotion-annotated dialogue corpusgenerates a prototype response that meets the contextual semantics. Secondly,the first-stage prototype is modified by a controllable emotion refiner withthe empathy hypothesis. Experimental results on the DailyDialog andEmpatheticDialogues datasets demonstrate that the proposed conversationaloutperforms the comparison models in emotion generation and maintains thesemantic performance in automatic and human evaluations.