CoMMA, a Large-scale Corpus of Multilingual Medieval Archives

Abstract

We present CoMMA, a large-scale corpus of medieval manuscripts produced through automatic text recognition. The corpus contains around 2.5b tokens drawn from more than 23,000 digitized manuscripts in Latin and Old French, harvested via IIIF. Unlike other resources, it is made of raw, non-normalized text enriched with layout analysis in various formats. We describe the pipeline used for large-scale acquisition and processing, and report quantitative and qualitative evaluations (average CER 9.7%). The resulting resource supports multiple use cases, from pretraining language models to corpus linguistic on historical languages and digital humanities applications.

Resources

Details