Projekt OCR-BW

Automatische Texterkennung von Handschriften

Abstract

After the digitization of historical documents, the next logical step is to enrich the digitized material with a searchable full text to further increase the accessibility of the texts and to enable new research questions. While many libraries already use various options for automatic text recognition of printed material, there is much higher reluctance to do so when it comes to manuscripts, since handwritten sources pose new challenges for automatic text recognition. With the help of machine learning, however, great progress has been made in the field of automatic handwritten text recognition in recent years, which libraries can not only use to make their own holdings more accessible, but also to establish themselves as a service partner for science. As part of the OCR-BW project (https://ocr-bw.bib.uni-mannheim.de/), since 2019 the transcription platforms Transkribus and, from 2021, eScriptorium have been systematically tested on selected corpora to generate automatic full texts for manuscripts. The results achieved during the project so far are very positive and show that automatic handwritten text recognition with a character error rate of less than 5 % is possible and can be expected. Full texts that have already been published have significantly increased the visibility and research interest in these materials. The project also aims to support science in the preparation and implementation of research projects. Examples ranging from medieval prayer books to large collections such as legal councils to expedition diaries of the 20th century will be used to show which results can be achieved with which resources.