[Paper Review] Generating Counterfactual Explanations with Natural Language

Lisa Anne Hendricks, Ronghang Hu+2 more|arXiv (Cornell University)|Jun 26, 2018

Explainable Artificial Intelligence (XAI)Computer Science53 citations

TL;DR

The paper proposes a method to generate counterfactual textual explanations for image classifications by identifying evidence discriminative for counter-classes, checking its presence in the image, and negating it to produce fluent counterfactual sentences. It evaluates on Caltech-UCSD Birds with automatic metrics for phrase errors and counterfactual-text impact.

ABSTRACT

Natural language explanations of deep neural network decisions provide an intuitive way for a AI agent to articulate a reasoning process. Current textual explanations learn to discuss class discriminative features in an image. However, it is also helpful to understand which attributes might change a classification decision if present in an image (e.g., "This is not a Scarlet Tanager because it does not have black wings.") We call such textual explanations counterfactual explanations, and propose an intuitive method to generate counterfactual explanations by inspecting which evidence in an input is missing, but might contribute to a different classification decision if present in the image. To demonstrate our method we consider a fine-grained image classification task in which we take as input an image and a counterfactual class and output text which explains why the image does not belong to a counterfactual class. We then analyze our generated counterfactual explanations both qualitatively and quantitatively using proposed automatic metrics.

Motivation & Objective

• Motivate and enable explanations that describe why an image does not belong to a counter-class, not just why it belongs to a class.
• Leverage semantic, non-image evidence to generate informative counterfactual statements.
• Develop an end-to-end pipeline that predicts counterfactual evidence, verifies its absence in the image, and generates fluent counterfactual text.
• Evaluate the quality and discriminativeness of counterfactual explanations with proposed metrics on a fine-grained dataset.

Proposed method

• Predict candidate counterfactual evidence from explanations of the counter-class using noun-phrase extraction from generated explanations.
• Verify whether the counterfactual evidence is present in the image using two evidence checkers: a Counterfactual: Classifier and a Counterfactual: Phrase-Critic.
• Negate the selected counterfactual phrases and compose cohesive sentences comparing to the counter-class (e.g., This is not a X because...).
• Use a rule-based negation system to form final counterfactual sentences and append to the base explanation.
• Optionally ground phrases in images via a retrieval-grounding model for localization to inform the phrase-critic scoring.
• Evaluate with phrase error and accuracy w/ counterfactual text on the Caltech-UCSD Birds dataset.

Experimental results

Research questions

• RQ1 Can counterfactual explanations improve interpretability by indicating attributes missing from an image that would change the class decision?
• RQ2 How accurately can a model predict and verify counterfactual evidence that is not present in the image?
• RQ3 Do counterfactual additions reduce the classifier’s ability to predict the correct class from explanations, indicating discriminativeness?
• RQ4 Which evidence checker (Classifier vs Phrase-Critic) better supports robust counterfactual text generation?

Key findings

• Both counterfactual models (CF: Classifier and CF: Phrase-Critic) outperform the baseline in reducing phrase errors in generated explanations.
• All models reduce the sentence-level accuracy when counterfactual text is added, indicating the text affects class-discriminative judgments.
• The Phrase-Critic model generally yields better phrase grounding performance and lower phrase error than the baseline and the classifier, suggesting improved localization of counterfactual attributes.
• Grounding-based approaches benefit from external data (e.g., Visual Genome) and phrase-level localization for more effective counterfactual evidence selection.
• The baseline remains strong in phrase error but is outperformed by the proposed counterfactual approaches in terms of reducing incorrect counterfactual mentions.
• Qualitative examples illustrate counterfactual explanations like “This is not a Bobolink because it does not have a yellow nape,” clarifying distinctions between similar birds.