공유저작물 자동 정보 추출을 위한 미세조정 BERT-BiLSTM-CRF NER 모델의 성능 및 안정성 분석 = Performance and Stability Analysis of a Fine-Tuned BERT-BiLSTM-CRF NER Model for Automated Information Extraction in Openly Licensed Works

다국어 초록 (Multilingual Abstract)

Korean legal documents pose challenges for information extraction due to complex layouts, Optical Character Recognition (OCR) noise, and agglutinative morphology. This paper proposes an automated Named-Entity Recognition(NER) pipeline that integrates Qwen-VL-based OCR, a Begin-Inside-Outside (B-I-O)-tagged training dataset, and fine-tuned BERT-family encoders with a BiLSTM-Conditional Random Field (CRF) decoder. We fine-tune mBERT, KLUE-RoBERTa-Large, and XLM-RoBERTa-Large under both Pure and BiLSTM-CRF settings, incorporating 30% OCR-style noise. A 5-폴드cross-validation demonstrates that CRF-enhanced models achieve more stable and structurally consistent predictions, with XLM-RoBERTa-Large-CRF reaching an average F1-score of 0.998. The results highlight a practical design for robust NER in noisy OCR environments.