LegNER: a domain-adapted transformer for legal named entity recognition and text anonymization

The increasing demand for scalable and privacy-preserving processing of legal documents has intensified the need for accurate Named Entity Recognition (NER) systems tailored to the legal domain. In this work, we introduce LegNER, a domain-adapted transformer model designed for both legal NER and text anonymization. The model is trained on a corpus of 1,542 manually annotated court cases and enriched with an extended legal vocabulary, enabling robust recognition of six critical entity types, including PERSON, ORGANIZATION, LAW, and CASE_REFERENCE. Built on BERT-base and enhanced through domain-specific pretraining and span-level supervision, LegNER consistently outperforms established legal NER baselines. Experimental results demonstrate significant gains in accuracy (99%), F1 score (over 99%), and inference efficiency (processing more than 12 documents per second), confirming both its precision and scalability. Beyond quantitative improvements, qualitative evaluation highlights LegNERs ability to generate coherent anonymized outputs, a crucial requirement for GDPR-compliant redaction and automated legal analytics. Taken together, these results establish LegNER as a reliable and effective solution for high-precision entity recognition and anonymization in compliance-sensitive legal workflows.