Transcribathon. Building a Sustainable Community of Practice around Historical Transcription and Handwritten Text Recognition

Abstract As the library at Norway’s oldest and largest university, we support research partnerships with academic staff, students, and faculty daily. Increasingly, we receive requests from researchers across disciplines for assistance with transcribing written historical documents. These range from Norwegian historical medical journals to 15th-century Tibetan sacred manuscripts, to early 20th-century field notes from archaeologists and anthropologists, to Arabic periodicals and early printed documents from the 19th century. We have learned that our patrons' needs are too diverse and, at the same time, too specialised for us to take on, so instead we offer assistance with learning to use tools like Transkribus for text recognition and with training new models. In addition to regular introductory tutorials and individual or project counselling with our subject specialists, we developed and tested a different mode of engagement: the Transcribathon. The Transcribathon, or Transcription Marathon, is based on the concept of a Hackathon, where a group of people who do not know each other from before get together for a full day at one of our venues and are tasked with transcribing or quality-checking machine transcriptions to either create enough Ground Truth for training an entirely new HTR model or publication-ready transcriptions of historical sources. A seasoned Transkribus user guides the events, and one or more subject specialists accompany participants to ensure everyone understands how text recognition works and how to navigate the Transkribus app on their own computers. Concrete goals for the Transcribathon are set prior to the event and communicated to participants, who will work individually or in small groups in 90-minute sessions, followed by 15-minute screen-free breaks. In my contribution to the Transkribus User Conference, I want to talk about two such events focusing on (1) specialised AI “Hard Work – Creating Ground Truth for a specialised 19th-century Arabic HTR Model in Transkribus“ and (2) the usability of general-purpose AI models “Worth the Hassle? Utilising general-purpose HTR models for multilingual historical letters from the Norwegian Observatory”. These examples will illustrate how we, as a cultural heritage and academic institution, navigate complex requests for research support with a unique and fun combination of training and guidance sessions that enable researchers to make informed decisions about the AI tools they use, the manual work they have to put in, and the expertise that is necessary to assess AI outputs for quality and reliability.