Towards Faithful Explanations for Text Classification with Robustness Improvement and Explanation Guided Training

TL;DR

REGEX addresses the fragility and misalignment of explanations in text classification by coupling robustness enhancement with explanation-guided training. It employs VAT and input gradient regularization to stabilize predictions, and uses Integrated Gradients to mask low-salience tokens while aligning model attention with attribution via KL-divergence terms. Across six datasets and multiple attribution methods, REGEX consistently improves faithfulness metrics (sufficiency and comprehensiveness) and robustness of explanations, while preserving or matching end-to-end task performance, including select-then-predict FRESH configurations. The work highlights a meaningful link between model robustness and explanation fidelity and outlines practical directions for building trustworthy NLP systems, while noting limitations in PLM diversity and evaluation scope.

Abstract

Feature attribution methods highlight the important input tokens as explanations to model predictions, which have been widely applied to deep neural networks towards trustworthy AI. However, recent works show that explanations provided by these methods face challenges of being faithful and robust. In this paper, we propose a method with Robustness improvement and Explanation Guided training towards more faithful EXplanations (REGEX) for text classification. First, we improve model robustness by input gradient regularization technique and virtual adversarial training. Secondly, we use salient ranking to mask noisy tokens and maximize the similarity between model attention and feature attribution, which can be seen as a self-training procedure without importing other external information. We conduct extensive experiments on six datasets with five attribution methods, and also evaluate the faithfulness in the out-of-domain setting. The results show that REGEX improves fidelity metrics of explanations in all settings and further achieves consistent gains based on two randomization tests. Moreover, we show that using highlight explanations produced by REGEX to train select-then-predict models results in comparable task performance to the end-to-end method.

Figures (4)

• Figure 1: Visualization of foored positive and foocyan negative highlights produced by post-hoc explanation methods (e.g., feature attribution). However, these explanations suffer from unfaithfulness problems (e.g., same model framework A and A' with different attributions) and can be further fooled by adversarial manipulation without changing model output Fragile (see §\ref{['sec:explan_robust']}).
• Figure 2: The overall framework of proposed REGEX method. REGEX consists of two components for robustness improvement and explanations guided training respectively. For latter, we iteratively mask input tokens with low attribution scores and then minimize the KL divergence between attention of masked input and feature attributions.
• Figure 3: Comparisons between different explanation guided training $\lambda_4$ on the SST dataset.
• Figure 4: Comparisons between different mask ratio $K$ on the SST dataset.