Knowledge-Grounded Self-Rationalization via Extractive and Natural Language Explanations

Explainability no longer requires sacrificing accuracy: The RExC framework achieves state-of-the-art task performance while simultaneously providing explanations, addressing a longstanding trade-off where models that explained their decisions typically performed worse than those optimized purely for accuracy.

Multiple explanation types serve different purposes: The framework provides both extractive rationales (highlighting key features) and natural language explanations, recognizing that different stakeholders may need different formats—quick feature summaries versus detailed reasoning—rather than forcing a one-size-fits-all approach.

Knowledge grounding improves explanation quality: By anchoring predictions and explanations in background knowledge, the model significantly improved the quality of both explanation types compared to previous methods, suggesting that AI systems perform better when they can reference established information rather than operating in isolation.

Faithful explanations show strong prediction-explanation alignment: Perturbation analysis demonstrated high association between the explanations generated and actual predictions, meaning the explanations genuinely reflect what the model is using to decide—a critical requirement for trustworthy AI systems.

Responsible AI requires both performance and interpretability: This work demonstrates that building trustworthy AI systems doesn't require choosing between accuracy and explainability; instead, thoughtful framework design can deliver both simultaneously, making it more feasible for organizations to deploy responsible AI in practice.