Can LLMs Credibly Transform the Creation of Panel Data from Diverse Historical Tables?∗
创建于 更新于
摘要
Multimodal LLMs enable low-cost digitization of heterogeneous historical tables by leveraging domain expertise via prompts, enabling cohesive panel data from diverse sources [page::0]. Using gold-standard data for evaluation, the pipeline achieves 95.4% exact cell matching at roughly 50x lower costs than outsourcing [page::0]. It also dramatically reduces layout parsing errors (61.4% to 0.35%), and attains 96.7% exact transcription of linked cells with a mean absolute percentage error of 0.7% [page::0]. In replication studies, the LLM pipeline aligns with human-based header categorization and yields econometric inferences that are largely indistinguishable from gold-standard results across eight models [page::0].
