NER-IPL: Indian Legal Prediction Dataset for Named Entity Recognition

Identifying Named Entities from unstructured text is difficult, especially for domain-specific data. Legal documents are usually very lengthy and highly unstructured. We can use two approaches for extracting the named legal entities from the legal documents: the Rule-based approach and the Machine-learning (ML) approach. This paper introduces NER-IPL, an Indian Legal Prediction Dataset for Named Entity Recognition. The dataset consists of 213481 sentences with 123193 annotated entities and 6198700 tokens. Different ML models take different encoding schemes to process the dataset; therefore, we use three different encoding schemes (BILOU, BOI, IOEBS) on the named entities for tagging, allowing the corpus to be trained on any machine learning model for automatic extraction of named entities. To validate our dataset, we have created a battery of baseline models to test the suitability of NER tasks using language models. All the different experiments with scores, detailed analysis, and the scope of improvements are elaborated in detail. Amongst the experimented baseline models, the InLegalBERT model gives the best F1 score of 0.67 on our dataset.