NEAR: Non-Supervised Explainability Architecture for Accurate Review-Based Collaborative Filtering