Improving Non-Autoregressive Response Generation by Distilling Posterior Information

NAR significantly improves generation efficiency for its fast parallel decoding mechanism; however, it suffers from the accuracy problem and logical coherence within the generation process. In order to enhance the logical coherence between the response and the question in response generation tasks, we model the topics of user questions to improve the stickiness of the reply content. We applied the human-machine dialogue system and evaluate it comparing with two other baseline methods. Experimental results show the high performance of our method in generating diverse and logical responses.