Explainable AI through Thematic Clustering and Contextual Visualization: Advancing Macro-Level Explainability in AFV Systems

Abstract

In this paper, we present a new framework for computing explainability in Automated Fact Verification models. Our approach uses a graph-based visualization mechanism based on thematic clustering, integrating Gaussian Mixture Models and an Expectation-Maximization algorithm, which are statistical methods for identifying patterns within data. These methods work together in our framework to deliver both local and global explainability. We present initial insights into applying our framework using a fact verification dataset consisting of claim evidence pairs, aiming to compute explainability for verification classifications at both the datapoint level (local) and across datapoints (global). We anticipate that integrating local and global perspectives will yield deeper insights into how individual claims fit within a cluster’s context, providing richer, context-sensitive explanations for fact verification tasks, where the context greatly influences the interpretation of individual facts.