"WHY IS THIS NEWS NEGATIVE?": AN ATTENTION-BASED VISUALIZATION TECHNIQUE FOR EXPLAINABLE SENTIMENT ANALYSIS. AN ANALYTICAL SURVEY AND CONCEPTUAL FRAMEWORK

Abstract

. The proliferation of Transformer-based architectures has marked a new epoch in Natural Language Processing, with sentiment analysis systems achieving unprecedented accuracy. This performance, however, is coupled with a profound lack of transparency, rendering these models as opaque "black boxes." This interpretability deficit is a significant barrier to their deployment in high-stakes domains like financial analysis and public policy, where the justification for a decision is as critical as the decision itself. This paper conducts a comprehensive survey of this problem and proffers a conceptual framework for an explainable sentiment analysis system. We propose leveraging the Transformer’s intrinsic self-attention mechanism to generate salient visualizations that elucidate the textual evidence driving a model’s classification. The paper examines the theoretical underpinnings of this method, engages with the pivotal "Attention is not Explanation" debate, and charts a course for future research. We present the architecture of a system engineered not merely to classify sentiment but to furnish a transparent rationale for its predictions, thereby fostering greater trust and accountability in advanced AI systems.