Towards Faithful Dialogues via Focus Learning

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4554–4566 July 9-14, 2023 ©2023 Association for Computational Linguistics Towards Faithful Dialogs via Focus Learning Yifan Deng1,2 Xingsheng Zhang1,2 ∗ Heyan Huang3 ∗ Yue Hu1,2 1 Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China 2 School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China 3 School of Computer Science and Technology, Beijing Institute of Technology, Beijing, China {dengyifan, zhangxingsheng, huyue}@iie.ac.cn hhy63@bit.edu.cn Abstract Maintaining faithfulness between responses and knowledge is an important research topic for building reliable knowledge-grounded dia- logue systems. Existing models heavily rely on the elaborate data engineering and increasing the model’s parameters ignoring to track the to- kens that significantly influence losses, which is decisive for the optimization direction of the model in each iteration. To address this issue, we propose Focus Learning (FocusL), a novel learning approach that adjusts the contribution of each token to the optimization direction by directly scaling the corresponding objective loss. Specifically, we first introduce a position- ing method by utilizing relevance distributions between knowledge and each response token to locate knowledge-aware tokens. Then, we further design a relevance-to-weight transfor- mation to provide dynamic token-level weights for adjusting the cross-entropy loss. Finally, we use the weighted loss to encourage the model to pay special attention to the knowledge uti- lization. Experimental results demonstrate that our method achieves the new state-of-the-art results and generates more reliable responses while maintaining training stability. 1 Introduction Although open-domain conversation systems can generate smooth and fluent responses with the help of large-scale pre-trained models (Raffel et al., 2020; Lewis et al., 2020), vacuous responses (Li et al., 2016) continue to be prevalent. To enrich the content of responses, an effective way is to intro- duce external knowledge (Dinan et al., 2019; Zhou et al., 2018). The knowledge-grounded model, however, frequently generates responses that ap- pear knowledgeable but are not derived from the given knowledge. This means that the correctness of the knowledge used in responses cannot be guar- anteed. As shown in Figure 1, the "Oklahoma" ∗Corresponding author Figure 1: The different learning focus (i.e., the to- kens that corresponding losses significantly influence the total objective loss) between general learning and focus learning. Original learning focus without guid- ance in general learning are fragmented with no rules to follow. Our methods make the model focus on the knowledge-aware tokens (i.e., tokens that have high semantic relevance to knowledge) to alleviate the hallucinations . in the response is not present in the given knowl- edge and relevant knowledge is unverifiable. This phenomenon is known as the hallucination (Dziri et al., 2022a) problem. Due to the inability to verify knowledge, hallucinations can mislead users and reduce the model’s credibility. Numerous methods have been developed to tackle the hallucination problem by knowledge graph (Kang et al., 2022; Dziri et al., 2021), con- trastive learning (Sun et al., 2022) or control code (Rashkin et al., 2021). These models enhance the …

Content truncated to 500 words