Towards Faithful Knowledge Graph Explanation Through Deep Alignment in Commonsense Question Answering

TL;DR

The LM-KG Fidelity metric is introduced to assess KG representation reliability and the LM-KG Distribution-aware Alignment algorithm is proposed to improve explanation faithfulness, highlighting the need to address distributional misalignment for reliable commonsense reasoning.

Abstract

The fusion of language models (LMs) and knowledge graphs (KGs) is widely used in commonsense question answering, but generating faithful explanations remains challenging. Current methods often overlook path decoding faithfulness, leading to divergence between graph encoder outputs and model predictions. We identify confounding effects and LM-KG misalignment as key factors causing spurious explanations. To address this, we introduce the LM-KG Fidelity metric to assess KG representation reliability and propose the LM-KG Distribution-aware Alignment (\textit{LKDA}) algorithm to improve explanation faithfulness. Without ground truth, we evaluate KG explanations using the proposed Fidelity-Sparsity Trade-off Curve. Experiments on CommonsenseQA and OpenBookQA show that LKDA significantly enhances explanation fidelity and model performance, highlighting the need to address distributional misalignment for reliable commonsense reasoning.

Figures (7)

• Figure 1: This figure depicts a class of models that integrate KG and LM for question answering. The training stage on the left side of the figure mainly includes LM, KG, and their interaction through a knowledge exchange fusion layer. The right side of the figure illustrates the post-hoc explanation results. Explanations extracted from the KG of models that produce the same correct answers can be inconsistent and unfaithful.
• Figure 2: Behavior of GNN model from the causality perspective in the form of Structural Equation Model. There are two possible causal paths can be found.
• Figure 3: This figure depicts the comprehensive structure of the fusion layer, through which the LM is deeply integrated with the KG. The components highlighted in pink signify the modules that exhibit a strong correlation with the LM. The purple dashed line denotes the specific segments that require LM detachment before the final prediction to keep GNN faithfulness.
• Figure 4: Illustration of our proposed new objective
• Figure 5: The bar charts compare the accuracy of the model on CommonsenseQA before and after LKDA training when the LM is detached. The models trained with LKDA are shown with a gray background.