RGTI:Response generation via templates integration for End to End dialog

Yuxin Zhang, Songyan Liu

Keywords: End-to-end dialogue systems, transformer, pointer-generate network

TL;DR: A new simple but efficient model for end-to-end dialogue

Abstract:

End-to-end models have achieved considerable success in task-oriented dialogue area, but suffer from the challenges of (a) poor semantic control, and (b) little interaction with auxiliary information. In this paper, we propose a novel yet simple end-to-end model for response generation via mixed templates, which can address above challenges.
In our model, we retrieval candidate responses which contain abundant syntactic and sequence information by dialogue semantic information related to dialogue history. Then, we exploit candidate response attention to get templates which should be mentioned in response. Our model can integrate multi template information to guide the decoder module how to generate response better. We show that our proposed model learns useful templates information, which improves the performance of "how to say" and "what to say" in response generation. Experiments on the large-scale Multiwoz dataset demonstrate the effectiveness of our proposed model, which attain the state-of-the-art performance.