LexiMus-NER: A Named Entity Recognition Dataset for Historical Spanish Musical Press (1918–1935)

Abstract

LexiMus-NER is the first Named Entity Recognition (NER) dataset designed specifically for historical Spanish musical press. Drawing on approximately 3,050,000 words from two early twentieth-century publications — El Sol (Madrid, 1918–1932) and Ondas (Madrid, 1925–1935) — the dataset defines four domain-specific entity types absent from general-purpose Spanish NER resources: COMPOSITOR (composer), INTERPRETE (instrumental performer or conductor), CANTANTE (singer), and AGRUPACION (musical ensemble). It comprises 11,697 annotated entities across 4,243 documents in CoNLL-style TSV/IOB2 format, accompanied by a curated gazetteer of 1,035 unique musical entities. A transformer-based model (BETO fine-tuned via spaCy, v9) achieves 91.9% entity-level detection precision on an independent 73,633-word evaluation corpus and above 90% human-validated classification precision across all four entity classes (n = 1,183). All resources are freely available on Zenodo (DOI: 10.5281/zenodo.19429405), Hugging Face, and GitHub under CC BY 4.0, produced within the LexiMus project ('Léxico y ontología de la música en español', PID2022-139589NB-C33), funded by the Spanish Ministry of Science and Innovation and the European Union.