Efficient OCR for Building a Diverse Digital History

Learn about groundbreaking approaches to Optical Character Recognition (OCR) in this 46-minute conference talk from Harvard's Big Data Conference. Professor Melissa Dell presents innovative methods for making digital archives more representative of diverse historical documents. Explore how treating OCR as a character-level image retrieval problem with contrastive learning leads to more efficient and extensible solutions compared to traditional sequence-to-sequence architectures. Discover how this approach requires fewer labeled examples and computational resources while maintaining accuracy, particularly beneficial for low-resource document collections. Follow along as Dell demonstrates practical applications, from supply chain network analysis to knowledge graph construction, and discusses how this technology enables greater community participation in digital preservation efforts. Gain insights into technical aspects including CRN architecture, object detection, data augmentation, and zero-shot performance across multiple languages, with special attention to Japanese character recognition results.