An Efficient Graph-Based Summarization Approach for Judicial Case Type Prediction Using BERT

The judiciary is one of the three pillars of democracy. Considering the increasing volume of legal cases, it has become vital to manage judicial resources judiciously. Determining the case type of legal documents is essential since it facilitates assigning cases to the right judges and court divisions. Recently, various NLP approaches have been devised to tackle this problem. However, the improvement in accuracy has been associated with rising model complexity. In this paper, an intuitive approach of pre-summarizing documents to save computing costs while retaining relevant information has been proposed. This enhances the computational efficiency of resource-intensive deep-learning models while dealing with large legal corpora. This has been followed by a BERT model fine-tuned on judicial case records for text classification. The proposed technique has been compared with a number of machine-learning and deep-learning techniques on the Supreme Court Judgment Prediction dataset. It yields encouraging results with an accuracy of 83.57% and over 50% less GPU usage due to parameters being reduced, indicating its potential to speed up legal document processing.