Leveraging Transfer Learning and Legal Named Entity Masking for Classification of Rhetorical Roles in Indian Legal Judgments

The lengthy and complex nature of Indian legal documents, such as judgment texts, makes manual information retrieval challenging, time-consuming and prone to errors. Legal document classification is an important task in the Indian Judicial system, where huge volumes of data are generated on a daily basis. There is a need for automating the classification of different sections within these judgments, in order to extract important legal insights and improve legal analysis. This research aims to address this by identifying the rhetorical roles of sentences in legal texts such as facts, arguments, precedents, and issues- by exploring a transfer learning-based approach on labelled Indian legal decision texts. A pre-trained language model fine-tuned for the Indian legal domain is used in combination with legal named masking for legal entities, to improve the model’s classification performance and generalization. The study also addresses the challenges related to the unstructured nature of the legal documents and class imbalance, where some labels/sections are cited more often than others. An F1-score of 72.54 is achieved through hyperparameter tuning with Optuna, using multiple features and TF-IDF based representations in combination with the pretrained InLegalBERT model. Additionally, explainable AI techniques are employed to provide transparency and trust in the predictions of the model, making the decision-making process clearer for legal professionals and improving the model’s applicability in the real-world scenarios.