A Legal-BERT Validated RAG Framework for Trustworthy AI-Assisted Legal Reasoning

Abstract

The integration of Artificial Intelligence (AI) into legal workflows presents both transformative potential and critical challenges, particularly regarding the reliability and legal validity of generated content. While large language models (LLMs) like GPT-4 demonstrate impressive fluency, they often fall short in domain-specific legal reasoning. This paper proposes a conceptual framework that enhances Retrieval-Augmented Generation (RAG) by incorporating Legal-BERT as an active validation layer. The architecture consists of five core components: document preprocessing using LayoutLMv3, summarization via BART, context retrieval with Dense Passage Retriever (DPR), generation through GPT-4, and legal validation using Legal-BERT. Central to the framework is a Hybrid Verification Loop, which iteratively refines outputs based on feedback from the validation stage. Designed for modularity and extensibility, the framework aims to produce legally compliant and interpretable outputs. While empirical evaluation is planned—initially focused on the Indian Penal Code—this work offers a foundation for future development of trustworthy legal AI systems, with potential for reinforcement learning-based enhancements.

Ähnliche Arbeiten