Fine-Grained Prototype-Based Interpretability for Operational Text Classification

Bowen Wei, Jinhao Pan, Ziwei Zhu

Keywords: Interpretability, Uncertainty, Text Classification

Abstract: We study interpretable, decision-centric NLP for operational settings that require accountability and robustness under uncertainty. We introduce \emph{ProtoLens}, a prototype-based model that produces fine-grained (sub-sentence) rationales aligned to semantically coherent prototypes, enabling principled integration with OR-style decision rules (e.g., cost- and risk-sensitive thresholds, audits, and overrides). Across text classification benchmarks, ProtoLens provides interpretable, human-aligned explanations while matching or exceeding competitive baselines.

Submission Number: 131