Crossing the Bifrost Towards an open access FAIR HTR model for Old Norse manuscripts

Résumé 0

This paper presents the initial results of a project devoted to creating the first-ever FAIR model for handwritten text recognition (HTR) of Old Norse-Icelandic manuscripts, accompanied by an open-access dataset of ground truths for manuscripts from the 13th and 14th centuries–the golden age of manuscript production in Iceland. Compared with other medieval languages written in Latin scripts, Old Norse-Icelandic does not benefit from the same amount of training data, which are crucial for developing strong HTR models. Therefore, we turn to an approach applied by scholars in other under-resourced fields, such as Arabic, to find sustainable solutions for the acquisition of texts. The goal of this paper is not only the presentation of another HTR model and its underlying dataset.1 By incrementally evaluating the accuracy of the model, we provide insights into the amount of data actuallyneeded to produce usable transcriptions. We provide reflections on the accessibility and influence of technology on research in smaller, often under-resourced disciplines, as well as on the position of individual researchers who are based outside large DH centres benefiting from state-of-the-art infrastructure. We also position our experiments with HTR in the broader context of research questions concerning Icelandic palaeography and scribal culture. One of the important points demonstrated by our paper is that all training and testing were done with Kraken (Kiessling 2019), run locally on a 2020 M1 MacBook Air (8 CPU, 16 GB). A comparable machine today can be purchased for around EUR 1.000, making it more accessible for many researchers than obtaining access to GPUs.