Generating Emotional Controllable Response Based on Multi-Task and Dual Attention Framework

Building an empathic conversation agent in open-domain is a key step towards affective computing and intelligent interactions. However, most current methods either focus on the consistency of content or the controllability of emotion and handling both factors are not yet properly solved. In this paper, we propose the multi-task and dual attentions (MTDA) framework for generating an emotional response. The MTDA framework decomposes the input utterance into the content layer and emotional layer, and then encodes and decodes them separately, which makes this end-to-end model more interpretable and controllable. A multi-task learning based encoder is employed in the MTDA framework, which can obtain the representation of the content and the emotion through unsupervised learning and supervised learning. A dual attention mechanism is adopted for decoding, which ensures that specific emotional responses are coherent with the content and the emotion of the input. We also combine the MTDA framework with state of the art generative models to train emotional generation systems. Extensive experiments show that our model can not only adapt to different target emotion goals but also generate coherent and informative responses.