Rhetorical Role Detection in Legal Judgements Using Zero-Shot Learning

In this paper, we address the problem of legal statement segmentation (or rhetorical role detection). Traditionally, this is handled by taking the expertise of lawyers and making them mark each and every statement as one of the many pre-defined classes. Naturally, this process is cumbersome and involves a lot of manual intervention. Zero-shot learning is a promising approach that could be one of the potential solutions to this labor-intensive problem. Therefore, in this paper, we apply zero-shot learning to the task of legal judgement segmentation. We try to remove the “human in the loop” and present a new potential direction in rhetorical role detection. To that end, we use BART to automatically classify various segments of a document into multiple classes. We propose a model that uses a pre-trained language model to generate embeddings for each document, which are then used to classify a legal sentence into one of the multiple classes. We evaluate our model on a dataset of legal documents consisting of manually marked statements. In particular, the dataset consists of 50 court case documents from the Indian Supreme Court. Through experimentation, we have found that the proposed gives a strong baseline that could act as a new direction in rhetorical role detection. Further, we also show that the model presented in this article can indeed pave the way for future work in legal analytics.