Explaining NLP Models via Minimal Contrastive Editing (MiCE)
This work addresses the lack of contrastive explanations in NLP model interpretability, which is important for developers to debug models and uncover dataset artifacts.
This paper introduces Minimal Contrastive Editing (MiCE), a method that generates minimal and fluent edits to input text to change an NLP model's prediction to a desired contrast case. The method was evaluated across sentiment classification, topic classification, and multiple-choice question answering, demonstrating its ability to produce human-consistent contrastive explanations.
Humans have been shown to give contrastive explanations, which explain why an observed event happened rather than some other counterfactual event (the contrast case). Despite the influential role that contrastivity plays in how humans explain, this property is largely missing from current methods for explaining NLP models. We present Minimal Contrastive Editing (MiCE), a method for producing contrastive explanations of model predictions in the form of edits to inputs that change model outputs to the contrast case. Our experiments across three tasks--binary sentiment classification, topic classification, and multiple-choice question answering--show that MiCE is able to produce edits that are not only contrastive, but also minimal and fluent, consistent with human contrastive edits. We demonstrate how MiCE edits can be used for two use cases in NLP system development--debugging incorrect model outputs and uncovering dataset artifacts--and thereby illustrate that producing contrastive explanations is a promising research direction for model interpretability.