Elastic Weight Removal for Faithful and Abstractive Dialogue Generation

TL;DR

This work tackles hallucination in document-grounded dialogue by introducing Elastic Weight Removal (EWR), an efficiency-preserving, parameter-level approach that weights the removal of undesirable behaviours by Fisher Information. By constructing anti-expert (hallucination) and expert (abstractive) task vectors and applying Fisher-weighted interpolation to the base model, EWR achieves systematic faithfulness gains with minimal cost to other metrics. The authors demonstrate robust improvements across WoW, DSTC9/11, and FaithDial, and show that combining EWR with CTRL yields state-of-the-art faithfulness in several settings, while also enabling controllable trade-offs between faithfulness and abstractiveness. They also provide extensive human evaluations and open-source code, highlighting practical impact and avenues for extending EWR to other domains and dimensions of desirable behaviour.

Abstract

Ideally, dialogue systems should generate responses that are faithful to the knowledge contained in relevant documents. However, many models generate hallucinated responses instead that contradict it or contain unverifiable information. To mitigate such undesirable behaviour, it has been proposed to fine-tune a `negative expert' on negative examples and subtract its parameters from those of a pre-trained model. However, intuitively, this does not take into account that some parameters are more responsible than others in causing hallucinations. Thus, we propose to weigh their individual importance via (an approximation of) the Fisher Information matrix, which measures the uncertainty of their estimate. We call this method Elastic Weight Removal (EWR). We evaluate our method -- using different variants of Flan-T5 as a backbone language model -- on multiple datasets for information-seeking dialogue generation and compare our method with state-of-the-art techniques for faithfulness, such as CTRL, Quark, DExperts, and Noisy Channel reranking. Extensive automatic and human evaluation shows that EWR systematically increases faithfulness at minor costs in terms of other metrics. However, we notice that only discouraging hallucinations may increase extractiveness, i.e. shallow copy-pasting of document spans, which can be undesirable. Hence, as a second main contribution, we show that our method can be extended to simultaneously discourage hallucinations and extractive responses. We publicly release the code for reproducing EWR and all baselines.

Figures (3)

• Figure 1: Constructed example of responses $u_{T+1}$ that are i) hallucinated (words contradicting the knowledge $\mathcal{K}$ in red); ii) faithful but not abstractive (longest copied $n$-gram in italic); and iii) both Faithful and Abstractive based on Wizard-of-Wikipedia dinan2018wizard.
• Figure 2: Metrics for EWR on top of Flan-T5$_\text{base}$ on the seen test split of WoW. (a) Varying the influence of the abstraction expert model gives control over the trade-off between faithfulness and abstractiveness, when fixing the scaling factor of the hallucination anti-expert. Dashed lines indicate baseline Flan-T5$_\text{base}$ performance and grey shading improvements over it in terms of both criteria. (b) Using only the hallucination anti-expert, varying the scaling factor again shows a decrease in Critic score but also an increase in Density and a slight decline in BLEU.
• Figure 3: Improvements in faithfulness (measured by Critic) tend to incur an increase in extractiveness (measured by LCS) on WoW.