Goal-Oriented Knowledge-Driven Neural Dialog Generation System

Abstract

In this paper, we propose a goal-oriented knowledge-driven neural dialog generation system, which leads the conversation based on a knowledge graph. During the conversation, the system has to actively integrate appropriate knowledge conditioned on current dialog state, and then generate coherent, fluent and meaningful responses. We use ERNIE as our backbone model, proposing a fine-tuning scheme to first pre-train on knowledge graph and dialog sequence, and then fine-tune to generate the next response. We extend multi-task learning in multi-turn dialog generation to improve consistency. We show that with well-designed transfer learning, ERNIE shows competitive performance on a knowledge-grounded dialog generation task. In the Baidu knowledge-driven dialog competition, our best single model achieved 4th in the automatic evaluation stage with 47.03 f1 score and 0.417/0.281 BLEU1/BLEU2 score, and ranked 1st in the final human evaluation stage, with descent topic completion performance(1.81/3) and highest coherence score(2.59/3).