Dialogue Generation with Skeleton Extraction and Causal Intervention

Retrieval and re-generate, which generates a natural language sentence through re-writing a retrieved sentence using a sequence-to-sequence generation structure, has been proven to be effective in dialogue response generation. However, the rewriting component tends to keep irrelevant information from the retrieval results, leading to low relevance in the re-written results. To address this problem, we propose a new framework (name), which extracts the skeleton from the retrieval results and uses it along with the query to generate responses. Specifically, we use the invariant learning method to capture the skeleton of the retrieval results. This solution employs causal intervention methods to identify noises in the retrieval results, enabling the generation component aware of what to keep and what to generate in the re-writing stage. Experimental results show that our approach improves the relevance and diversity of the generated results compared to the existing generative models.