Human-grounded Evaluations of Explanation Methods for Text Classification

TL;DR

This paper evaluates various explanation methods for CNN-based text classification through human-centered studies, revealing their strengths and limitations for different interpretability purposes.

Contribution

It provides a comprehensive human-grounded assessment of multiple explanation techniques tailored for text classification models, highlighting their practical utility.

Findings

Different explanation methods excel at specific interpretability tasks.

Some methods effectively reveal model behavior, others justify predictions.

Explanation quality varies significantly across methods.

Abstract

Due to the black-box nature of deep learning models, methods for explaining the models' results are crucial to gain trust from humans and support collaboration between AIs and humans. In this paper, we consider several model-agnostic and model-specific explanation methods for CNNs for text classification and conduct three human-grounded evaluations, focusing on different purposes of explanations: (1) revealing model behavior, (2) justifying model predictions, and (3) helping humans investigate uncertain predictions. The results highlight dissimilar qualities of the various explanation methods we consider and show the degree to which these methods could serve for each purpose.