Crossing the Divide: Designing Layers of Explainability

In the era of deep learning, the opaque nature of sophisticated models often stands at odds with the growing demand for transparency and explainability in Artificial Intelligence. This paper introduces a novel approach to text classification that emphasizes explainability without significantly compromising performance. We propose a modular framework to distill and aggregate information in a manner conducive to human interpretation. At the core of our methodology is the premise that features extracted at the finest granularity are inherently explainable and reliable; compared with methods whose explanation is on word-level importance, this layered aggregation of low-level features allows us to trace a clearer decision trail of the model's decision-making process. Our results demonstrate this approach yields effective explanations with a marginal reduction in accuracy, presenting a compelling trade-off for applications where understandability is paramount.