Eliciting Knowledge from Large Pre-Trained Models for Unsupervised Knowledge-Grounded Conversation

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10551 - 10564 December 7-11, 2022 ©2022 Association for Computational Linguistics Eliciting Knowledge from Large Pre-Trained Models for Unsupervised Knowledge-Grounded Conversation Yanyang Li, Jianqiao Zhao, Michael R. Lyu, Liwei Wang∗ Department of Computer Science and Engineering, The Chinese University of Hong Kong {yyli21,jqzhao,lyu,lwwang}@cse.cuhk.edu.hk Abstract Recent advances in large-scale pre-training pro- vide large models with the potential to learn knowledge from the raw text. It is thus nat- ural to ask whether it is possible to leverage these large models as knowledge bases for downstream tasks. In this work, we answer the aforementioned question in unsupervised knowledge-grounded conversation. We explore various methods that best elicit knowledge from large models. Our human study indicates that, though hallucinations exist, large models post the unique advantage of being able to output common sense and summarize facts that can- not be directly retrieved from the search engine. To better exploit such generated knowledge in dialogue generation, we treat the generated knowledge as a noisy knowledge source and propose the posterior-based reweighing as well as the noisy training strategy. Empirical results on two benchmarks show advantages over the state-of-the-art methods. 1 Introduction Knowledge-grounded conversation (Dinan et al., 2019; Moghe et al., 2018) is the task where the model could reply to a dialogue history based on extra knowledge. Compared to standard con- versational modeling, this extra knowledge pre- vents the model from generating generic and non- informative responses (Li et al., 2016). Typically, at each turn of the conversation, a pool of knowledge candidates will be retrieved from a knowledge base like unstructured documents (e.g., Wikipedia) (Di- nan et al., 2019) or a structured knowledge graph (Dziri et al., 2021). The model then learns to select the most related knowledge from this pool, in an unsupervised manner, to generate its response. However, constructing and maintaining knowl- edge bases are time-consuming and expensive. Re- cent studies have shown that large pre-trained mod- ∗Corresponding author. els are capable of grasping knowledge from un- supervised text corpora and memorizing facts to their weights (Petroni et al., 2019; Roberts et al., 2020; Lewis et al., 2021; Wang et al., 2021; Liu et al., 2022a). These large models can even per- form reasoning implicitly (Wei et al., 2022). In light of this remarkable capacity of large models, we explore the possibility of leveraging large mod- els as a new knowledge source for unsupervised knowledge-grounded conversation. In this work, we first investigate the quality of knowledge generated by large models. We exam- ine which tuning method, including the conven- tional fine-tuning (Devlin et al., 2019; Zhang et al., 2021) and the recently proposed prefix-tuning (Li and Liang, 2021), best prompts knowledge from large models for a given dialogue history. We then design a human evaluation protocol and conduct an extensive quality assessment of the generated knowledge. Despite some extent of hallucinations (plausible statements with factual errors) persist (Maynez et al., 2020), large models can mostly generate related and correct …

Content truncated to 500 words