Explanation-guided learning improves the transparency of aspect-based sentiment analysis by coupling predictions with supporting evidence. However, the explanatory signals used for supervision are often unstable under small input perturbations and semantically incomplete at the token level, which limits their effectiveness for training guidance. To address this problem, we propose IEG, an internal–external explanation-guided framework that integrates three complementary components. First, IEG performs importance-guided data augmentation by editing low-importance tokens to create semantics-preserving variants. Second, it applies explanation consistency regularization through sufficiency and comprehensiveness constraints to stabilize internal rationales. Third, it aligns gradient-based evidence with phrase-level rationales generated by an LLM, thereby introducing semantically richer external supervision. Mathematically, IEG is formulated as a composite regularized empirical-risk objective that couples classification loss with sufficiency, comprehensiveness, and external alignment penalties over aspect-conditioned inputs. Experiments on three benchmark datasets show that IEG consistently outperforms strong explanation-guided baselines, with gains of up to 1.77 accuracy points and 2.69 macro-F1 points, while also improving rationale-oriented evaluation as higher comprehensiveness and lower sufficiency.
Lin et al. (Wed,) studied this question.