OBOE: an Explainable Text Classification Framework

Abstract

Explainable Artificial Intelligence (XAI) has recently gained visibility as one of the main topics of Artificial Intelligence research due to, among others, the need to provide a meaningful justification of the reasons behind the decision of black-box algorithms. Current approaches are based on model agnostic or ad-hoc solutions and, although there are frameworks that define workflows to generate meaningful explanations, a text classification framework that provides such explanations considering the different ingredients involved in the classification process (data, model, explanations, and users) is still missing. With the intention of covering this research gap, in this paper we present a text classification framework called OBOE (explanatiOns Based On concEpts), in which such ingredients play an active role to open the black-box. OBOE defines different components whose implementation can be customized and, thus, explanations are adapted to specific contexts. We also provide a tailored implementation to show the customization capability of OBOE. Additionally, we performed (a) a validation of the implemented framework to evaluate the performance using different corpora and (b) a user-based evaluation of the explanations provided by OBOE. The latter evaluation shows that the explanations generated in natural language express the reason for the classification results in a way that is comprehensible to non-technical users.