Automating MARC Records from Turkish Title Pages: A Performance and Efficiency Comparison of Fine-Tuned BERT vs. VLMs

The automation of bibliographic cataloging, particularly the generation of MARC records from title page images, presents a significant challenge for libraries, especially for under-resourced languages such as Turkish. This study introduces an end-to-end pipeline for this task and provides a rigorous comparative analysis of two distinct AI approaches: a domain-specific, fine-tuned BERT model (Biblio-BERTurk-NER-Base) and general-purpose Vision Language Models (VLMs), including Gemini 2.5 Flash and Qwen2.5-VL-7B-Instruct. Biblio-BERTurk-NER-Base was trained on a custom-annotated dataset of 484 Turkish book title pages, with this study also investigating the impact of augmenting the training data with 300 synthetic samples in a follow-up experiment (Biblio-BERTurk-NER-Extended). The models were evaluated not only on standard NER metrics (F1-score, Precision, and Recall) but also on resource consumption, including inference latency, energy usage, and cost. Results indicate that Biblio-BERTurk-NER-Base achieved the highest performance with a macro F1-score of 0.666, outperforming the leading VLM, Gemini 2.5 Flash (F1: 0.533). Counter-intuitively, the addition of synthetic data degraded performance, with Biblio-BERTurk-NER-Extended reducing its F1-score to 0.536. Furthermore, the analysis of resource consumption reveals that Biblio-BERTurk-NER-Base is overwhelmingly more efficient, demonstrating significantly lower cost and carbon footprint per sample compared to the VLM alternatives. This study provides a practical blueprint for libraries handling non-English collections and offers empirical evidence that for specialized, text-based metadata extraction, a smaller, domain-adapted model presents a more performant, cost-effective, and sustainable solution than larger, general-purpose models.