Retrieval-Augmented Language Models for Automated Legal Document Analysis and Clause Classification

Abstract

The rapid growth of digital legal documents has increased the demand for automated methods capable of extracting, organizing, and interpreting complex contractual information. Traditional natural language processing approaches frequently struggle with long-range dependencies, domain-specific terminology, and variations in legal document structures. This study introduces a retrieval-augmented language modeling framework for automated contract clause classification and evidence-supported legal information extraction. The proposed approach combines semantic document retrieval with instruction-tuned language models to improve contextual understanding and reduce unsupported predictions. The illustrative experimental dataset consists of approximately 186,000 contract clauses extracted from 24,500 legal agreements across 17 commercial contract categories. The proposed framework achieves a macro-averaged F1-score of 0.918 and a clause extraction accuracy of 94.1%, compared with 0.856 and 88.3%, respectively, for a conventional transformer-based baseline. Furthermore, evidence citation precision reaches 92.7%, while unsupported statement generation decreases by 34.6%. These findings illustrate the potential of retrieval-augmented artificial intelligence to improve the efficiency, traceability, and consistency of legal document processing.