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

TL;DR

This paper investigates how large pre-trained models can be used as sources of knowledge for unsupervised knowledge-grounded conversation, proposing methods to extract and utilize their generated knowledge effectively.

Contribution

It introduces novel techniques for eliciting and leveraging knowledge from large models in dialogue systems, improving over existing methods.

Findings

Large models can produce common sense and factual summaries.

Proposed methods outperform state-of-the-art on benchmark datasets.

Generated knowledge is noisy but valuable for dialogue generation.

Abstract

Recent advances in large-scale pre-training provide large models with the potential to learn knowledge from the raw text. It is thus natural 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 cannot 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.