Multi Task Learning for Legal Judgement Prediction

Abstract

In Legal Judgement Prediction (LJP), the primary aim is to anticipate the judgment outcome by analyzing the factual description of criminal or civil cases, a field gaining increasing attention within legal research. Typically, a LJP scenario include three core subtasks: predicting applica ble law articles, charges, and penalty terms. In real-world scenarios, both predicting charges and legal articles necessitate multi-class classification combined with elements of multi-label learn ing. However, current methods frequently reduce these tasks to single-label learning paradigms for multi-class categorization. Furthermore, these methods typically overlook the utilization of pertinent keywords abundant in legal documents. The critical analysis of violated laws and related articles to formulate a final judgement is a very time-consuming task. Hence, an au tomated prediction of judgement can fulfill huge amount of repetitive tasks in minimal time. Therefore, we propose a model for the task of multi-class multi-label prediction formalizing the dependencies among the sub tasks to improve the multi-task prediction efficiency using a framework titled as Multi Task Learning for Legal Judgement Prediction. This framework lever ages a multi-layered attention network, machine learning and transformer models that integrates legal keywords, fostering a comprehensive understanding of multiple LJP subtasks. Analyzing and implementing the law heavily relies on legal texts, necessitating a thorough examination of violated laws and relevant articles to reach a final judgment. Therefore, this research assist judges in decision making process based on the valid references and an automated judgement prediction simplifies the whole case proceedings.