Automatically Classify each sentence from Indian Legal Judgements to its Semantic Role

In the past decade, automation has gained traction in Indian law. To perform many downstream tasks like judgment classification, judgment prediction, Named Entity Recognition (NER), etc., digitally understanding Indian legal case documents/judgments is the most important task. In legal judgement documents, sentences perform different semantic functions. Many models have sought in recent years to use contextual awareness between sentence sequences using Conditional Random Field (CRF) over manually handcrafted features, Hierarchical BiLSTM CRFs, etc., to assimilate the interrelationship between consecutive sentences. We propose a model for Semantic Role Labelling (SRL) of each sentence in Indian legal judgements. Our model out-performs previous models by using Named Entity replacement followed by Graph Attention (GAT) mechanism to generate context-aware sentence embeddings for each judgement sentence. An efficient SRL model makes downstream task of summarizing each semantic segment of judgement, more efficient and speed-up judgement analysis/study by law practitioners to deliver justice to all. A multi-class classifier neural network receives these context-aware embeddings and classifies them into one of thirteen semantic categories. The task is challenging since legal judgement documents are not properly structured. Also, the semantics roles may be subjective.