TextFocus: Assessing the Faithfulness of Feature Attribution Methods Explanations in Natural Language Processing

Abstract

<p>Among the existing eXplainable AI (XAI) approaches, Feature Attribution methods are a popular option due to their interpretable nature. However, each method leads to a different solution, thusintroducing uncertainty regarding their reliability and coherence with respect to the underlying model. Thiswork introduces TextFocus, a metric for evaluating the faithfulness of Feature Attribution methods for NaturalLanguage Processing (NLP) tasks involving classification. To address the absence of ground truth explanationsfor such methods, we introduce the concept of textual mosaics. A mosaic is composed of a combination ofsentences belonging to different classes, which provides an implicit ground truth for attribution. The accuracyof explanations can be then evaluated by comparing feature attribution scores with the known class labels inthe mosaic. The performance of six feature attribution methods is systematically compared on three sentenceclassification tasks by using TextFocus, with Integrated Gradients being the best overall method in terms offaithfulness and computational requirements. The proposed methodology fills a gap in NLP evaluation, byproviding an objective way to assess Feature Attribution methods while finding their optimal parameters.</p>