This project focuses on identifying 14 different types of legal entities within the Indian Judicial Dataset. We employ a multi-pronged approach, leveraging cutting-edge models like DeBERTa and ELECTRA, as well as utilizing Zero-shot and One-shot Named Entity Recognition (NER).

We used the Indian Legal NER Dataset (Prathamesh Kalamkar, 2022), which is a corpus of 46,545 annotated legal named entities mapped to 14 legal entity types. The dataset is extracted from Preamble and Judgement documents of the Indian judicial system.

• Fine-tuning Advanced BERT Models: DeBERTa and ELECTRA
• Zero-shot and One-shot Learning: Transforming multi-class token classification into binary token classification
• Prompt Engineering: Used GPT 3.5 and experimented with zero-shot, one-shot, and few-shot learning scenarios.
• Baseline: Spacy NER component pipeline

Detailed report along with results are summarized here.

This is a group project done as a part of a graduate UCSD course Statistical Natural Language Processing.