Topic-level knowledge sub-graphs for multi-turn dialogue generation

Previous multi-turn dialogue approaches based on global Knowledge Graphs (KGs) still suffer from generic, uncontrollable, and incoherent responses generation. Most of them neither consider the local topic-level semantic information of KGs nor effectively merge the information of long dialogue contexts and KGs into the dialogue generation. To tackle these issues, we propose a Topic-level Knowledge-aware Dialogue Generation model to capture context-aware topic-level knowledge information. Our method thus accounts for topic-coherence, fluency, and diversity of generated responses. Specifically, we first decompose the given KG into a set of topic-level sub-graphs, with each sub-graph capturing a semantic component of the input KG. Furthermore, we design a Topic-level Sub-graphs Attention Network to calculate the comprehensive representation of both sub-graphs and previous turns of dialogue utterances, which then decoded with the current turn into a response. By using sub-graphs, our model is able to attend to different topical components of the KG and enhance the topic-coherence. We perform extensive experiments on two datasets of DuRecDial and KdConv to demonstrate the effectiveness of our model. The experimental results demonstrate that our model outperforms existing strong baselines.