PR-LIME: Two-Stage Perturbation Refinement for Phrase-Level Text Model Explanations

Existing perturbation-based explanation methods such as LIME for text tasks provide only token-level importance scores and neglect the collective contribution of contiguous phrases. To address this limitation, we propose PR-LIME, a two-stage perturbation refinement framework that yields coherent phrase-level explanations. In the first stage, we adopt LIME’s large-scale binary-mask sampling to obtain token-level priors. In the second stage, we introduce a BiGRU-based mask generator, which refines perturbation samples under the joint constraints of fidelity, continuity prior, and KL-divergence loss to learn coherent phrase-level masks. Finally, ridge regression integrates samples from both stages to produce contribution weights for each token or phrase. On the SST-2 text classification benchmark, PR-LIME achieves a 22.3% relative improvement over the original LIME in fidelity metrics; qualitative visualizations further demonstrate the superiority of phrase-level explanations in both continuity and fidelity.