Deep Learning-Based Semantic Segmentation for Legal Texts: Unveiling Rhetorical Roles in Legal Case Documents

Abstract

The swift rise of digitization in legal documentation has opened doors for artificial intelligence to revolutionize various tasks within the legal domain. Among these tasks is the segmentation of legal documents using rhetorical labels. This process, known as rhetorical role labeling, involves assigning labels (such as Final Judgment, Argument, Fact, etc.) to sentences within a legal case document. This task can be down streamed to various major legal analytics problems such as summarization of legal documents, readability of lengthy case documents, document similarity estimation, etc. The mentioned task of semantic segmentation of documents via labels is challenging as the legal documents are lengthy, unstructured and the labels are subjective in nature. Various previous works on automatic rhetorical role labeling was carried out using methods like conditional random fields with handcrafted features, etc. This research focuses on analyzing case documents from two different legal systems: the High Court of Kerala and the High Court of Justice in the United Kingdom. Through rigorous experimentation with a range of deep learning models, this study highlights the robustness and efficacy of deep learning methods in accurately labeling rhetorical roles within legal texts. Additionally, comprehensive annotation of legal case documents from the UK and analysis of inter-annotator agreement are conducted. The overarching objective of this research is to design systems that facilitate a deeper comprehension of the organizational structure inherent in legal case documents.

Keywords

• legal document segmentation
• rhetorical role labeling
• meta embedding
• hierarchical bi-lstm crf,bert