This National Science Foundation Project Grant award of $595,690 provides funding from July 1, 2022 through June 30, 2025 to develop new techniques for unsupervised Islamic manuscript transcription via lacunae reconstruction. The award is made through the NSF's Computer and Information Science and Engineering program (CFDA 47.070) to support investigator-initiated research and education in all areas of computing, communications, and information science and engineering.
Specifically, the University of Maryland, College Park will develop a novel unsupervised learning framework to transcribe Islamic manuscripts without relying on large amounts of supervised training data. This is necessary given the wide variation in scribal hands across Islamic texts and the scarcity of expert-transcribed data available for supervised training of modern neural networks for handwritten text recognition. The framework centers on pretraining an encoder to reconstruct masked regions of unlabeled manuscript images, implicitly learning to encode discrete symbols. If successful, this new technique has the potential to unlock vast archives of premodern Islamic written cultural production for advanced scholarly analysis through accurate transcription of entire manuscript collections.