Legal Chunking: Evaluating Methods for Effective Legal Text Retrieval

Andrea Filippo Ferraris, Davide Audrito, Giovanni Siragusa, Alessandro Piovano

Abstract: Legal document retrieval is heavily influenced by how documents are segmented, or “chunked,” for processing within Retrieval-Augmented Generation (RAG) systems. This paper investigates the effectiveness of three automated chunking techniques — Simple Text Splitting, Recursive Text Splitting using regular expressions (regex), and Semantic Chunking — within the legal domain, using the General Data Protection Regulation (GDPR) as a testbed. The chunking methods were evaluated based on their semantic relevance to a set of seventeen legal questions and their corresponding relevant sections, with performance measured using cosine similarity metrics. Results show that none of the methods consistently produced high semantic relevance on an individual chunk level: Git Hub link.

External IDs:doi:10.3233/faia241255