Explainable Transformers in Natural Language Processing

Transformer-based models have revolutionized natural language processing while becoming increasingly opaque in their decision-making processes. As these systems deploy in high-stakes domains such as healthcare, legal decision-making, and content moderation, their lack of transparency poses critical challenges, particularly where explainability is legally mandated or ethically required. This dissertation introduces ExpNet, a novel framework that learns to generate human-interpretable explanations from transformer attention mechanisms through supervised learning. Unlike existing approaches that apply fixed mathematical heuristics or rely on computationally expensive perturbation methods, ExpNet treats explanation generation as a learning problem. The framework extracts attention features and processes them through a lightweight neural network trained on human-annotated rationales, discovering optimal mappings from attention patterns to token importance scores. ExpNet demonstrates strong cross-task generalization capability through rigorous evaluation on diverse natural language processing tasks including sentiment analysis, grammatical acceptability judgment, and hate speech detection. When trained on tasks from certain domains and evaluated on entirely different tasks, ExpNet consistently outperforms thirteen baseline methods spanning model-agnostic, gradient-based, and attention-based approaches. This generalization achieved without task-specific retraining suggests that transformer attention encodes universal linguistic patterns, enabling the framework to explain models across a broad range of applications beyond its training domains. Quantitative results show ExpNet achieves substantially higher alignment with human rationales than all baselines across varied linguistic phenomena. The framework generates explanations orders of magnitude faster than perturbation-based methods, making it suitable for real-time deployment. This combination of superior explanation quality and computational efficiency addresses the critical gap between interpretability research and practical production requirements. The dissertation develops extensive analysis of explanation evaluation challenges in natural language processing. By learning from human reasoning, generalizing across diverse tasks, and maintaining computational efficiency, ExpNet provides a pathway toward transparent transformer-based systems suitable for high-stakes applications where explainability is imperative.