Position Debiasing Fine-Tuning for Causal Perception in Long-Term Dialogue

TL;DR

This work addresses position bias in large language models when handling long-term dialogues, which hampers true causal perception of historical utterances. It introduces CPD, a model-free framework that (1) extracts causally relevant utterances via perturbation and a local-position awareness module, and (2) applies causal-perception finetuning using invariant risk minimization and treatment-effect-based objectives with a carefully designed sampling strategy. The causal view models dialogue with $D$, $R$, $C$, and $S$, using $\text{TE}(u_i) = f(D) - f(D \setminus u_i)$ to identify influential turns and enforce invariance across perturbations through $\mathcal{L} = \mathcal{L}_{Pred} + \alpha \mathcal{L}_{IRM} + \beta \mathcal{L}_{MTE}$. Experiments on ESConv and MSC show that CPD consistently outperforms baselines on automatic metrics and human judgments, demonstrating improved causal perception and reduced position bias, with ablations confirming the importance of IRM, MTE, and the sampling scheme.

Abstract

The core of the dialogue system is to generate relevant, informative, and human-like responses based on extensive dialogue history. Recently, dialogue generation domain has seen mainstream adoption of large language models (LLMs), due to its powerful capability in generating utterances. However, there is a natural deficiency for such models, that is, inherent position bias, which may lead them to pay more attention to the nearby utterances instead of causally relevant ones, resulting in generating irrelevant and generic responses in long-term dialogue. To alleviate such problem, in this paper, we propose a novel method, named Causal Perception long-term Dialogue framework (CPD), which employs perturbation-based causal variable discovery method to extract casually relevant utterances from the dialogue history and enhances model causal perception during fine-tuning. Specifically, a local-position awareness method is proposed in CPD for inter-sentence position correlation elimination, which helps models extract causally relevant utterances based on perturbations. Then, a casual-perception fine-tuning strategy is also proposed, to enhance the capability of discovering the causal invariant factors, by differently perturbing causally relevant and non-casually relevant ones for response generation. Experimental results on two datasets prove that our proposed method can effectively alleviate the position bias for multiple LLMs and achieve significant progress compared with existing baselines.

Figures (7)

• Figure 1: Position bias in large language models (Llama2-7B-chat). Dotted boxes mark relevant utterances. The strip on the right shows the average attention of each turn in the dialogue. Darker colors represent higher attention values.
• Figure 2: Causal view of response generation, where the solid line represents the causal relationship between two variables, and the dotted line represents the probabilistic dependencies.
• Figure 3: Llama2-7B-chat's ability to identify causally relevant utterances in the CGDIALOG dataset.
• Figure 4: The Framework of our proposed method.
• Figure 5: Our method (with Llama2-7B-chat as backbone)’s ability to identify multi causally relevant utterances. The solid blue line depicts the treatment effect as the number of perturbed causally relevant utterances is equal to the x-axis value. The red dashed line represents the treatment effect when the number of perturbed causally relevant utterances matches the x-axis value at the starting point of the blue line and the remains are uncausally relevant utterances.