A Multilingual Hybrid News Recommendation Framework for Educational Web Portals

University web portals increasingly serve as vital platforms for academic information sharing, yet effective news recommendation in resource-constrained, multilingual environments remains challenging due to limited labeled data, sparse user profiles, and linguistic diversity. This study presents a modular hybrid news recommendation framework tailored for educational web portals in low-resource settings. The approach integrates lexical methods, specifically Term Frequency–Inverse Document Frequency (TF–IDF) and Best Match 25 (BM25), with semantic retrieval based on Sentence-BERT (SBERT), combined with unsupervised clustering for topical diversification and a fuzzy-logic fusion layer to integrate heterogeneous similarity signals. A publicly available multilingual dataset of 1,389 university news articles was collected via a custom crawler, and a Flask-based API was implemented for real-time recommendation. Evaluation relies on an automatic hybrid ground truth generated by fusing SBERT, TF–IDF, and BM25 signals. On the ground truth subset, the hybrid model attains Precision@5 = 0.96 and NDCG@5 = 0.945, outperforming SBERT (Precision@5 = 0.93; NDCG@5 = 0.859), with improvements shown to be statistically significant (paired t-test on NDCG@5, p < 1e-5). Clustering enhances thematic diversity (entropy 1.697 vs. 0.032), reducing concentration on repeated announcements. Multilingual experiments demonstrate consistently high precision across Arabic, Kurdish, and English but reveal substantially lower recall for underrepresented languages, highlighting dataset imbalance and representation challenges. Fusion weights were tuned on a validation split to balance precision and recall while mitigating the dominance of any single signal across languages and content types. The proposed framework provides an interpretable and extensible solution for multilingual academic news recommendation in scenarios where interaction data are scarce, offering a practical foundation for future work on language-aware preprocessing, human validation of labels, and supervised re-ranking.