Coarse-to-Fine Response Generation for Document Grounded Conversations

Generating fluent and informative responses is crucial for dialogue systems. Most existing studies introduce documents as additional document to improve dialogue generation performance. However, understanding complex and multi-topic documents as well as extracting useful information related to the dialogue context from them is challenging. To solve this problem, we propose a Coarse-to-Fine Response Generation (CFRG) model with an encoder-decoder for document-grounded conversations,where the encoder is used to aware the themes of the dialogue context, and the decoder parse the semantics of replies. Specifically, the CFRG extracts coarse-grained features guided by the response from the context and documents and then uses them as input of decoder. It further interacts with the current sequence in the cross-attention layer to generate fine-grained features, on which a non-linear transformation function is trained to map to a word space. The experimental results on datasets CMU_DoG and Wizard of WikiPedia show that our model demonstrates superior performance compared to the majority of other models and achieves the best or comparable results.