Goodhart's Law Applies to NLP's Explanation Benchmarks

TL;DR

This paper critically examines NLP explanation benchmarks, revealing they can be manipulated without changing model predictions, which questions their reliability for guiding explainability research.

Contribution

It demonstrates that existing explanation metrics like ERASER and EVAL-X can be arbitrarily inflated, exposing their limitations and prompting a reassessment of evaluation standards.

Findings

Metrics can be inflated without changing model predictions

Current benchmarks are vulnerable to simple manipulations

Results question the reliability of explanation metrics

Abstract

Despite the rising popularity of saliency-based explanations, the research community remains at an impasse, facing doubts concerning their purpose, efficacy, and tendency to contradict each other. Seeking to unite the community's efforts around common goals, several recent works have proposed evaluation metrics. In this paper, we critically examine two sets of metrics: the ERASER metrics (comprehensiveness and sufficiency) and the EVAL-X metrics, focusing our inquiry on natural language processing. First, we show that we can inflate a model's comprehensiveness and sufficiency scores dramatically without altering its predictions or explanations on in-distribution test inputs. Our strategy exploits the tendency for extracted explanations and their complements to be "out-of-support" relative to each other and in-distribution inputs. Next, we demonstrate that the EVAL-X metrics can be inflated arbitrarily by a simple method that encodes the label, even though EVAL-X is precisely motivated to address such exploits. Our results raise doubts about the ability of current metrics to guide explainability research, underscoring the need for a broader reassessment of what precisely these metrics are intended to capture.