Fine-tuning SBERT for Semantic Research Title Classification in Trilingual University Repository

Recommendation systems are essential for automatically surfacing relevant content from large datasets, reducing search time, and facilitating discovery. In academia, content-based recommendation systems are beneficial when only brief titles are available and multilingual text is standard. Universities in the Kurdistan Regional Government currently lack a centralized research repository, with records scattered across different institutions and often manually maintained. This makes it difficult for students and faculty to find related topics, potential supervisors, or cross-disciplinary connections. This paper presents a trilingual (English, Arabic, and Kurdish) recommendation system for academic research titles. Three key contributions are made: (1) the creation of the first integrated dataset of 4,257 research titles from Sulaimani Polytechnic University publicly available; (2) the development of a web-based platform for semantic search and title-level recommendations to support research discovery and student–supervisor matching; and (3) an evaluation between Sentence-BERT models—all-MiniLM-L6-v2 and paraphrase-multilingual-MiniLM-L12-v2—before and after fine-tuning with a domain-specific taxonomy and cosine embedding loss. Performance is assessed using Precision@5, Mean Reciprocal Rank, and NDCG@5 with expert-annotated relevance judgments for 20 query titles. Fine-tuning resulted in performance improvements, with paraphrase-multilingual-MiniLM-L12-v2 achieving Precision@5 of 0.94 and NDCG@5 of 0.991. The English-only model also showed improvements, Precision@5: 0.79→0.82; NDCG@5: 0.885→0.922.