Comparative analysis of explainable AI techniques SHAP and LIME for deep neural network classification

Rapid advances in large language models have enabled deep neural networks (DNNs) to achieve exceptional performance across a range of complex real‑world tasks. Despite these successes, the opaque or “black‑box” nature of DNNs continues to pose challenges for interpreting and communicating their decision‑making processes to end users. This study examines the use of explainable AI (XAI) techniques within the domain of natural language processing (NLP) to enhance understanding of model behavior. Leveraging the Bidirectional Encoder Representations from Transformers (BERT) architecture, we generate local feature‑attribution explanations using two widely adopted XAI frameworks: LIME and SHAP. To evaluate whether these methods offer complementary insights, we perform statistical correlation analysis on their explanation outputs across multiple datasets and classification scenarios. The results indicate that, although the explanations produced by LIME and SHAP frequently diverge in feature rankings and attribution magnitudes, their combined use provides richer interpretability than either tool alone. In particular, the complementary perspectives offered by the two methods can deepen end‑user understanding of model decisions and strengthen trust in DNN‑based NLP systems. By demonstrating how LIME and SHAP can jointly illuminate the internal behavior of a state‑of‑the‑art language model, this work contributes to ongoing efforts to bridge the gap between high‑performing AI systems and human comprehension of their outputs.