EAGLE: Enhance Target-Oriented Dialogs by Global Planning and Topic Flow Integration

In this study, we propose a novel model EAGLE for target-oriented dialogue generation. Without relying on any knowledge graphs, our method integrates the global planning strategy in both topic path generation and response generation given the initial and target topics. EAGLE comprises three components: a topic path sampling strategy, a topic flow generator, and a global planner. Our approach confers a number of advantages: EAGLE is robust to the target that has never appeared in the training data set and able to plan the topic flow globally. The topic path sampling strategy samples topic paths based on two predefined rules and use the sampled paths to train the topic path generator. The topic flow generator then applies a non-autoregressive method to generate intermediate topics that link the initial and target topics smoothly. In addition, the global planner is a response generator that generates a response based on the future topic sequence and conversation history, enabling it to plan how to transition to future topics smoothly. Our experimental results demonstrate that EAGLE produces more coherent responses and smoother transitions than state-of-the-art baselines, with an overall success rate improvement of approximately 25% and an average smoothness score improvement of 10% in both offline and human evaluations.