CS2S-NCG: Cycle-Seq2Seq Model for Open Domain Neural Conversation Generation

摘要

Neural conversation generation model has attracted more and more attention in open domain conversation generation. Even though NCG-based conversation generation methods obtain promising results, they tend to generate short, generic and uninformative responses for most inputs. Moreover, training NCG models heavily relies on a large amount of quality paired dialogue corpus which tends to be expensive and difficult. In this paper, a unified Cycle-Seq2Seq model is proposed for fully data-driven Neural Conversation Generation (CS2S-NCG) in open domain. CS2S-NCG has ability to simultaneously learn the relevance of paired dialogue corpus, utilize feedback about the quality of the two generation models and reduce dependency on the amount of paired training dataset. Meanwhile, a joint learning strategy supports enhanced generation by better usage of massive unpaired dialogue corpus. An alternating Joint-EM algorithm is developed to solve this joint model. A series of experiments are conducted to evaluate CS2S-NCG on two automated metrics, as well as with human evaluation. The results demonstrate that CS2S-NCG significantly improves the relevance and diversity of generation results, compared to traditional neural conversation generation models. Furthermore, quantitative results also reveal that conversation generation quality are enhanced by leveraging unpaired dialogue corpus for not only CS2S-NCG but also traditional NCG.