An Interpretable Hybrid Ensemble Model for Paraphrase Detection With SHAP-Based Explanations
Paraphrase detection is a fundamental task in natural language processing, typically addressed using high-performing deep neural models that lack interpretability. Although transformer-based and hybrid systems attain high precision, their decision-making procedures are still opaque, which limits their dependability in practical applications. The explainable hybrid framework for paraphrase detection presented in this study combines transformer-based models, semantic similarity models, and recurrent neural networks into a single voting architecture. Token-level attribution of model predictions is provided by SHapley Additive exPlanations (SHAP), which are integrated to improve transparency. Furthermore, a semantic-aware optimization method is used to steer model behavior in the direction of significant similarity patterns. The proposed method is evaluated on the QQP and MRPC, and PAWS benchmark datasets, achieving accuracies of 85.4%, 88.24%, and 94.36% on the QQP, MRPC, and PAWS benchmark datasets respectively. demonstrating strong performance and generalization capability. The results show that the framework provides a balanced trade-off between accuracy and interpretability for paraphrase detection systems.