Measuring Association Between Labels and Free-Text Rationales
This addresses the need for reliable interpretability in NLP models for reasoning tasks, though it is incremental as it builds on existing joint models.
The paper tackled the problem of ensuring faithfulness in free-text natural language rationales for interpretable NLP, demonstrating that state-of-the-art T5-based joint models exhibit label-rationale association through robustness equivalence and feature importance agreement on commonsense QA and NLI tasks.
In interpretable NLP, we require faithful rationales that reflect the model's decision-making process for an explained instance. While prior work focuses on extractive rationales (a subset of the input words), we investigate their less-studied counterpart: free-text natural language rationales. We demonstrate that pipelines, existing models for faithful extractive rationalization on information-extraction style tasks, do not extend as reliably to "reasoning" tasks requiring free-text rationales. We turn to models that jointly predict and rationalize, a class of widely used high-performance models for free-text rationalization whose faithfulness is not yet established. We define label-rationale association as a necessary property for faithfulness: the internal mechanisms of the model producing the label and the rationale must be meaningfully correlated. We propose two measurements to test this property: robustness equivalence and feature importance agreement. We find that state-of-the-art T5-based joint models exhibit both properties for rationalizing commonsense question-answering and natural language inference, indicating their potential for producing faithful free-text rationales.