Generating Emotional Responses with DialoGPT-Based Multi-task Learning

Emotion is an essential element for a high quality response, and generating emotional responses is of great significance to a social dialogue system. Traditionally, training an emotional dialogue system needs a large-scale dialogue corpus with emotion labels, which is too expensive for mannual annotation while automatic labeling quality is not guaranteed. Multi-task learning provides a way to learn an emotional response generator through information sharing with emotion recognition tasks. This paper proposes a multi-task learning architecture based on DialoGPT, incorporating response generation with several emotion recognition tasks of different emotion granularities, including both singlelabel and multi-label classification. Experiments from both automatic evaluation and human evaluation show that the proposed models can generate emotional responses of high quality, outperforming all baseline models in most metrics.