Ekstraksi Informasi pada Dokumen Putusan Pengadilan Negeri Surabaya Perkara Pidana Narkotika Menggunakan Metode Named Entity Recognition Berbasis Legal-BERT

The large number of narcotics court decisions produced each year in Indonesia creates challenges for manual legal document analysis. This study aims to develop an automated information extraction approach using a Named Entity Recognition (NER) model based on Legal-BERT to identify important entities in narcotics criminal court decisions. The dataset consisted of 323 court decisions collected from the Supreme Court’s Decision Directory and manually annotated using the BIO scheme for four entity types: defendant names, evidence, fines, and imprisonment duration. The Legal-BERT model was fine-tuned for token classification and evaluated using precision, recall, and F1-score metrics. The experimental results show that the model achieved strong overall performance, with micro F1-score of 0.817 and macro F1-score of 0.838, indicating good balance between detection accuracy and completeness. Per-entity evaluation demonstrates that imprisonment duration achieved the best performance, while evidence entities remain the most challenging due to their varied and complex structure. These findings confirm that Legal-BERT is effective for legal NER tasks and can support automated information extraction from Indonesian court decisions, although further improvements are needed to enhance precision and entity boundary detection.