Identifying Social Class and Networks from US Legislative Journals using LLMs

Historical documents, such as nineteenth-century American legislative journals, contain vast amounts of structured information, yet they remain largely inaccessible for quantitative analysis due to non-machine-readable formats and optical character recognition (OCR) noise. This paper presents a scalable, Large Language Model (LLM)-based framework designed to extract structured roll-call data from large corpora of noisy, image-derived text. To overcome LLM context window constraints, the proposed pipeline employs a four-stage "hierarchical narrowing" strategy—indexing, classifying, assembling, and extracting—coupled with internal cross-referencing (matching extracted names to aggregate vote counts) for automated validation. Using the South Carolina House of Representatives (1858–1882) as a demonstration case, the pipeline achieves an extraction accuracy exceeding 95%. The resulting dataset facilitates the construction of legislative co-voting networks, allowing for the analysis of latent factional structures across the dramatic political ruptures of the antebellum, Civil War, and Reconstruction eras. Ultimately, this approach offers a highly generalizable template for recovering sparse, structured data from historical text at scale.