A Cause-Effect Look at Alleviating Hallucination of Knowledge-grounded Dialogue Generation

TL;DR

This work addresses hallucination in knowledge-grounded dialogue generation by introducing a Structural Causal Model (SCM) that links dialogue history $D$, retrieved knowledge $K$, dialogue context $C$, and response $R$. It identifies the Total Direct Effect ($TDE$) as a key metric and proposes a training-free counterfactual decoding method that subtracts the counterfactual knowledge influence $f_R(C_{d^{*},k})$ from the actual generation $f_R(C_{d,k})$, thereby reducing hallucinations while preserving dialogue quality. The implementation uses a dual decoding scheme with token-level adjustment $g_i = \arg\max\left(p(g_i) - \lambda(i) p(\hat{g}_i)\right)$ and $\lambda(i)=\alpha^{i-1}$ to realize the $TDE$-driven control at inference. Experiments on the Chinese KDConv dataset show reduced hallucination and robust performance across models like GLM and CTXL, demonstrating a lightweight, adaptable approach to building more faithful knowledge-grounded dialogue systems.

Abstract

Empowered by the large-scale pretrained language models, existing dialogue systems have demonstrated impressive performance conducting fluent and natural-sounding conversations. However, they are still plagued by the hallucination problem, causing unpredictable factual errors in the generated responses. Recently, knowledge-grounded dialogue generation models, that intentionally invoke external knowledge resources to more informative responses, are also proven to be effective in reducing hallucination. Following the idea of getting high-quality knowledge, a few efforts have achieved pretty good performance on this issue. As some inevitable knowledge noises may also lead to hallucinations, it is emergent to investigate the reason and future directions for building noise-tolerant methods in KGD tasks. In this paper, we analyze the causal story behind this problem with counterfactual reasoning methods. Based on the causal effect analysis, we propose a possible solution for alleviating the hallucination in KGD by exploiting the dialogue-knowledge interaction. Experimental results of our example implementation show that this method can reduce hallucination without disrupting other dialogue performance, while keeping adaptive to different generation models. We hope our efforts can support and call for more attention to developing lightweight techniques towards robust and trusty dialogue systems.

Figures (5)

• Figure 1: An example of the hallucination of knowledge-grounded dialogue. The figure presents three scenarios where the retrieved knowledge is perfectly matched, mismatched and with some redundant information. The texts with shading are dialogue-relevant knowledge. Green ones means the accurate contents. Red ones are the possible causes and effects of hallucination.
• Figure 2: (a) Casual graph for KGD; (b) Comparison between conventional KGD (left) and counterfactual KGD (right). White nodes are at the value $D = d$ and $K = k$ while gray one is $D = d^{*}$. Node $C$, $R$ are counterfactual.
• Figure 3: An illustration of counterfactual decoding. Anti-Hallu. is abbreviated from Anti-Hallucination. This mechanism encourages to improve the TDE of the dialogue $D$. The final generated response is more effected by dialogue contexts.
• Figure 4: Comparison of the distribution of fine-grained hallucination labels of all comparison methods. The horizontal axis represents the percentage of hallucination labels.
• Figure 5: The comparison of the distribution on user involvement distribution of GLM w/ and w/o AH mechanism. We show the number of conversation rounds (y-axis) and the total time involved in the conversation (x-axis).