Does It Make Sense to Explain a Black Box With Another Black Box?

TL;DR

This paper compares transparent and opaque counterfactual explanation methods in NLP, finding that opaque approaches often add unnecessary complexity without improving performance in tasks like fake news detection and sentiment analysis.

Contribution

It provides a comparative analysis of two main families of counterfactual explanation methods in NLP and questions the necessity of using black-box explanations for black-box models.

Findings

Opaque methods do not significantly outperform transparent ones.

Opaque approaches add complexity without performance gains.

Using black-box explanations for black-box models may be unnecessary.

Abstract

Although counterfactual explanations are a popular approach to explain ML black-box classifiers, they are less widespread in NLP. Most methods find those explanations by iteratively perturbing the target document until it is classified differently by the black box. We identify two main families of counterfactual explanation methods in the literature, namely, (a) \emph{transparent} methods that perturb the target by adding, removing, or replacing words, and (b) \emph{opaque} approaches that project the target document into a latent, non-interpretable space where the perturbation is carried out subsequently. This article offers a comparative study of the performance of these two families of methods on three classical NLP tasks. Our empirical evidence shows that opaque approaches can be an overkill for downstream applications such as fake news detection or sentiment analysis since they add an additional level of complexity with no significant performance gain. These observations motivate our discussion, which raises the question of whether it makes sense to explain a black box using another black box.