LEGAL JUDGEMENT PREDICTION USING LARGE LANGUAGE MODELS FOR INDIAN COURT CASES

Abstract

The application of Natural Language Processing to legal text understanding has gained substantial momentum in recent years, driven by the need to make judicial systems more accessible, efficient, and analytically transparent. Court Judgment Prediction (CJP) — the computational task of determining whether an appeal in a given court case will be accepted or rejected — sits at the intersection of legal reasoning and machine learning, and represents one of the most practically consequential problems in Legal Artificial Intelligence. The predominant approaches in this space rely on supervised fine-tuning of large transformer based language models, methods that, while effective, impose considerable computational and data requirements that render them inaccessible for many academic and low-resource settings. This dissertation introduces RR-RAG (Rhetorical Role-aware Retrieval-Augmented Gen eration), a lightweight and computationally efficient framework for Court Judgment Predic tion in the Indian legal domain that deliberately avoids expensive supervised fine-tuning. The central insight motivating the framework is that different sections of a legal judgment carry asymmetric predictive value: precedent citations and judicial ratio sections encode the reasoning chain most directly relevant to outcome, whereas purely descriptive factual sec tions introduce noise that can impair prediction quality. RR-RAG exploits this structural asymmetry by segmenting each judgment into rhetorical roles — FACT, ISSUE, PRECE DENT, RATIO/ANALYSIS, and RULING — retaining only the precedent and ratio segments as condensed legal representations, and indexing these representations in a FAISS-based se mantic memory built from sentence embeddings produced by BAAI/bge-base-en-v1.5. At inference time, the top-k semantically similar prior judgments are retrieved from the index and supplied as structured few-shot examples to the instruction-tuned language model Qwen2.5-3B-Instruct, which performs precedent-guided legal reasoning before emitting a binary accept/reject prediction. Experiments conducted on the CJPE subset of the IL TUR dataset demonstrate that rhetorical-role-conditioned retrieval outperforms full-text retrieval on accuracy, macro F1, and recall, while achieving competitive precision. The framework requires no gradient updates, operates entirely with pre-trained weights, and can be deployed on a single consumer-grade GPU.