Leakage-Adjusted Simulatability: Can Models Generate Non-Trivial Explanations of Their Behavior in Natural Language?
This work addresses the challenge of ensuring model-generated explanations are faithful to actual behavior rather than just plausible, which is crucial for improving transparency and trust in AI systems, though it is incremental in refining evaluation metrics.
The paper tackles the problem of evaluating natural language explanations generated by models for their behavior, arguing that current evaluations based on similarity to human explanations are insufficient for assessing faithfulness. They introduce a leakage-adjusted simulatability (LAS) metric to measure how well explanations help predict model outputs while controlling for leakage, and show that one rationalizing method achieves roughly human-level LAS scores on CoS-E and e-SNLI datasets.
Data collection for natural language (NL) understanding tasks has increasingly included human explanations alongside data points, allowing past works to introduce models that both perform a task and generate NL explanations for their outputs. Yet to date, model-generated explanations have been evaluated on the basis of surface-level similarities to human explanations, both through automatic metrics like BLEU and human evaluations. We argue that these evaluations are insufficient, since they fail to indicate whether explanations support actual model behavior (faithfulness), rather than simply match what a human would say (plausibility). In this work, we address the problem of evaluating explanations from the model simulatability perspective. Our contributions are as follows: (1) We introduce a leakage-adjusted simulatability (LAS) metric for evaluating NL explanations, which measures how well explanations help an observer predict a model's output, while controlling for how explanations can directly leak the output. We use a model as a proxy for a human observer, and validate this choice with two human subject experiments. (2) Using the CoS-E and e-SNLI datasets, we evaluate two existing generative graphical models and two new approaches; one rationalizing method we introduce achieves roughly human-level LAS scores. (3) Lastly, we frame explanation generation as a multi-agent game and optimize explanations for simulatability while penalizing label leakage, which can improve LAS scores. We provide code for the experiments in this paper at https://github.com/peterbhase/LAS-NL-Explanations