A tri-level context-aware framework for explainable SMS spam detection using semantic embeddings and fuzzy uncertainty modeling

SMS spam and smishing are persistent threats to cyber security, but numerous detectors are based on a single-view representation or function as a black box. This study proposes a leakage-safe, tri-level hybrid framework for explainable SMS spam detection that combines: (1) statistical TF-IDF features, (2) semantic embeddings derived from BiLSTM, (3) fuzzy linguistic indicators conveying graded uncertainty cues (keyword density, urgency, promotional intent, suspicious patterns). The generalized bell membership functions with train-fold P 95 -based calibration are used to map fuzzy scores to five linguistic states. A concatenated heterogeneous feature vector is classified through an XGBoost ensemble and local, global, and block-wise explanations are generated by aggregating attributions over the statistical, semantic, and fuzzy feature blocks, using TreeSHAP. The proposed model is evaluated using a strict repeated stratified cross-validation protocol (5 folds × 3 repeats) on the primary SMS Spam Collection benchmark, obtaining 98.76% accuracy (mean ± std over 15 held-out evaluations), and 95.22% spam-class F1-score, outperforming the strongest single-block baseline (Wilcoxon p = 6.5 × 10 − 4 ). A leave-one-block-out ablation study shows that primary predictive performance is attributed to the BiLSTM embeddings, while fuzzy cues provide valuable complementary gains (a 1.07% F1-score drop when removed). Cross-dataset validation on a modern Mendeley SMS phishing corpus achieves an accuracy of 98.57% and a 96.16% spam-class F1-score, indicating the framework's robustness and architectural adaptability to current distributions.