Named Entity Recognition (NER) is a core task in Natural Language Processing and Digital Humanities, yet its performance remains uneven for low-resourced languages and domain-specific genres. While English benefits from literary NER resources such as LitBank, Romanian lacks annotated datasets tailored to narrative fiction. This study addresses this gap by investigating domain-specific NER for Romanian fairytales, a challenging genre characterized by magical beings, symbolic entities, and fictional spaces that exceed standard annotation schemes.

We introduce FairyNERo, a manually curated gold-standard corpus of 101 open-access children’s fairytales (approximately 74,000 words). To capture genre-specific features, we developed a domain-adapted ontology extending the traditional PER, LOC, and TIME tags used in CoNLL. A hybrid human-in-the-loop workflow combined automatic pre-annotation with manual correction via a custom interface. Inter-annotator agreement reached 94.6%, indicating strong consistency.

We compared three approaches: (1) applying the general Romanian NER model RONEC, (2) fine-tuning a transformer model based on BERT-base-ro, and (3) prompting Microsoft Copilot in zero-shot and few-shot settings. The off-the-shelf model performed poorly due to domain mismatch. Fine-tuned BERT achieved the best results, with a strict span-level F1 score of 0.92 under CoNLL evaluation, demonstrating the effectiveness of domain adaptation. Large language model prompting outperformed the baseline but remained inferior to supervised fine-tuning.

Our findings highlight that ontology design and manual annotation are interpretive acts encoding genre knowledge essential for high-quality NER. By integrating ontology-driven annotation and transformer fine-tuning, this study provides the first Romanian fairytale NER resource and a replicable framework for domain-specific NER in low-resourced languages.