Contextual Knowledge Learning for Dialogue Generation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 7822–7839 July 9-14, 2023 ©2023 Association for Computational Linguistics Contextual Knowledge Learning for Dialogue Generation Wen Zheng1, Natasa Milic-Frayling1,2, Ke Zhou1,3 1University of Nottingham, 2Qatar Computing Research Institute, 3Nokia Bell Labs {wen.zheng, natasa.milic-frayling, ke.zhou}@nottingham.ac.uk Abstract Incorporating conversational context and knowledge into dialogue generation models has been essential for improving the quality of the generated responses. The context, comprising utterances from previous dialogue exchanges, is used as a source of content for response gen- eration and as a means of selecting external knowledge. However, to avoid introducing ir- relevant content, it is key to enable fine-grained scoring of context and knowledge. In this pa- per, we present a novel approach to context and knowledge weighting as an integral part of model training. We guide the model train- ing through a Contextual Knowledge Learn- ing (CKL) process which involves Latent Vec- tors for context and knowledge, respectively. CKL Latent Vectors capture the relationship between context, knowledge, and responses through weak supervision and enable differen- tial weighting of context utterances and knowl- edge sentences during the training process. Ex- periments with two standard datasets and hu- man evaluation demonstrate that CKL leads to a significant improvement compared with the performance of six strong baseline models and shows robustness with regard to reduced sizes of training sets. 1 Introduction Dialogue generation is concerned with conversa- tional settings where participants take turns and the task is to generate a response to previous ut- terances. In order to generate relevant responses, prior research explored the use of the conversa- tional history, i.e., the utterances already exchanged between the participants, as a context for the new response. It also investigated the use of external knowledge that may be a good source of content and not necessarily present in the conversational history (Ghazvininejad et al., 2018; Zheng and Zhou, 2019; Liu et al., 2021a,b; Prabhumoye et al., *Qatar Computing Research Institute, Doha, Qatar Figure 1: Example showing utterances of participants A and B, the scored knowledge, and the target response. The knowledge sentence (1) is deemed the best knowl- edge for response generation by our proposed CKL model. The best context segments for retrieving the best knowledge are colored Brown; the best context seg- ments for generating response are colored Blue and the best knowledge segments for the response are Purple. 2021a). However, it has been shown that adding knowledge indiscriminately, can hurt performance. Thus, the context has been used to select the best knowledge for the response generation. Since the context itself consists of multiple utter- ances, the same concern applies: not all the prior utterances are equally useful for generating the response. Therefore, the context needs to be eval- uated for its importance in relation to generating the response and identifying the relevant knowl- edge, separately. In Figure 1, we show an example from the Wizard of Wikipedia (test seen) dataset, illustrating that the best context segments for re- …

Content truncated to 500 words