FRAME: Evaluating Rationale-Label Consistency Metrics for Free-Text Rationales
This work addresses the reliability of evaluation metrics for free-text rationales in AI, which is crucial for improving interpretability in language models, though it is incremental as it builds on existing rationale consistency concepts.
The paper tackled the problem of evaluating metrics for free-text rationales in language models, which are prone to hallucination, by proposing FRAME, a framework based on three axioms to assess rationale-label consistency metrics. The result showed that existing metrics fail to satisfy all axioms, and a new non-pretraining metric outperformed baselines, achieving high scores on two axioms and competitive performance on the third.
Following how humans communicate, free-text rationales aim to use natural language to explain neural language model (LM) behavior. However, free-text rationales' unconstrained nature makes them prone to hallucination, so it is important to have metrics for free-text rationale quality. Existing free-text rationale metrics measure how consistent the rationale is with the LM's predicted label, but there is no protocol for assessing such metrics' reliability. Thus, we propose FRAME, a framework for evaluating rationale-label consistency (RLC) metrics for free-text rationales. FRAME is based on three axioms: (1) good metrics should yield highest scores for reference rationales, which maximize RLC by construction; (2) good metrics should be appropriately sensitive to semantic perturbation of rationales; and (3) good metrics should be robust to variation in the LM's task performance. Across three text classification datasets, we show that existing RLC metrics cannot satisfy all three FRAME axioms, since they are implemented via model pretraining which muddles the metric's signal. Then, we introduce a non-pretraining RLC metric that greatly outperforms baselines on (1) and (3), while performing competitively on (2). Finally, we discuss the limitations of using RLC to evaluate free-text rationales.