🤖 AI Summary
This work addresses the challenge of effectively integrating human-provided natural language explanations into end-to-end classifier training—a limitation of prior methods that struggle to incorporate explanations as direct, differentiable supervision. We propose Reflective-Net, an attention-based dual-channel neural network that, for the first time, models natural language explanations as differentiable supervisory signals, jointly optimizing classification accuracy and explanation alignment. Our approach employs a multi-task loss to enable explanation-driven end-to-end training. Evaluated on reasoning-intensive NLP benchmarks (e.g., e-SNLI), Reflective-Net achieves significant improvements in classification accuracy, explanation fidelity, zero-shot transferability, and counterfactual robustness. The core contribution is a novel joint learning paradigm that bridges prediction and explanation within a fully differentiable, trainable framework—advancing both model performance and interpretability.
