Elastic Weight Removal for Faithful and Abstractive Dialogue Generation

Anonymous

TL;DR: We show how model parameter interpolation can be used to reduce the amount of hallucinations in dialogue models.

Abstract: Generating factual responses is a crucial requirement for dialogue systems. To promote more factual responses, a common strategy is to ground their responses in relevant documents that inform response generation. However, common dialogue models still often hallucinate information that was not contained in these documents and is therefore unfaithful. In this work, we propose to alleviate such hallucinations by 'subtracting' the parameters of a model trained to hallucinate from a dialogue response generation model in order to 'negate' the contribution of such hallucinated examples from it. Extensive automatic and human evaluation shows favourable results when compared to state-of-the-art methods that combine the distributions of multiple models, such as DExperts (Liu et al., 2021), and others that change the training procedure, such as Quark (Lu et al., 2022a). Finally, we show how we can not only reduce hallucinations but also discourage extractive responses, which are often a consequence of reducing hallucinations by encouraging copy-pasting of document spans. We will publicly release our code for reproducibility and facilitating further research.

Paper Type: long

Research Area: Dialogue and Interactive Systems

Contribution Types: NLP engineering experiment

Languages Studied: English