Few Shot Rationale Generation using Self-Training with Dual Teachers

Self-rationalizing AI can work with limited labeled data: The paper demonstrates that models can generate both predictions and explanations using few-shot learning (minimal labeled examples), which reduces the expensive and time-consuming process of having humans write explanations for every training example.

Unlabeled data significantly improves explanation quality: By using a "dual-teacher" approach that learns from both labeled and unlabeled data through self-training, the model generates more faithful and trustworthy rationales—meaning the explanations actually reflect the reasoning behind the AI's decisions rather than being post-hoc justifications.

Explanations must be tied to actual predictions: The paper introduces a new technique called Masked Label Regularization that ensures generated explanations are genuinely conditioned on what the model predicted, addressing a key concern in responsible AI where explanations can sometimes be misleading or disconnected from the actual decision-making process.

This approach scales explanation generation: By reducing dependence on human-written rationales at scale, this method makes it more practical for organizations to deploy self-rationalizing AI systems across different tasks and domains, supporting the broader goal of building more transparent and trustworthy AI applications.