Beyond Embeddings: Comparing Language-Aware Sparse Representations and Fine-Tuned Multilingual Transformer Classifiers for Multi-label Subject Classification in Institutional Repositories

Abstract

Institutional repositories need scalable support for multilingual subject metadata assignment to long documents. On 19,710 SEDICI items and 37 labels, we compare 231 sparse/fixed-embedding configurations and two fine-tuned DistilUSE/LaBSE classifiers under separate threshold-selection, validation, and test partitions. Language-aware BM25 full text with LinearSVC achieved held-out Macro-F1 = 0.7522 (95% CI 0.7372–0.7643), Micro-F1 = 0.7685 (0.7567–0.7797), and Recall@3 = 0.9320; the LaBSE-initialized classifier reached Macro-F1 = 0.7069. A post-hoc BM25–LaBSE matched-exposure control, validation-only over three seeds with no test access, preserved the aggregate BM25 advantage. Strong sparse baselines therefore remain competitive under the evaluated Transformer protocols.