Indian Legal-Specialized NER Model Released

The en_legal_ner_ind_trf v0.1 model, fine-tuned on InLegalBERT using approximately 34,700 silver-annotated data points, has been released.

This model was developed to overcome the performance limitations of OpenNyAI's model, which was previously the only Indian legal NER model, and it shows stable performance even on OCR text data from before 1990.

Key performance metrics:

• CASE_CITATION: 97.76% F1 (+17 points over the existing baseline)
• PROVISION: 96.35% F1
• STATUTE: 91.94% F1
• Overall micro avg F1: 78.67%

Data annotation pipeline:

• Regex: Extraction of case citations, statutes, and provisions
• Metadata projection: Mapping of judge, petitioner, and respondent information
• Transformer NER: Extraction of lawyers, courts, organizations, place names, etc.
• Gazetteer: Supplementing information using 858 central statutes data

This model is distributed under the Apache-2.0 license.