On Curating HTR Training Datasets for Romanian Language with use of Transcribathon Tool
TL;DR
The workflow applied to prepare datasets of Romanian historical documents for training language-specific HTR models or for enhancing the language coverage of general purpose models is presented and a preliminary evaluation of the effort required to create HTR training sets is presented.
Abstract
Fulltext digitisation of historical manuscripts is an important activity in Digital Archives. The impressive development of handwritten text recognition technology provides the required technical support for mass (fulltext) digitization of archival materials. The quality of the automatically recognised text has become high for a couple of intensively used languages like English, German, French, etc. However, this is not the case for under-represented languages such as Romanian, for which, according to our current level of knowledge, there is currently no HTR model openly available. While Transkribus provides more than 300 public models, none of them support the Romanian language in Latin script. There are two experimental models available in Transkribus (https://www.transkribus.org/), one for Romanian in Cyrillic script and one for the transitional script, both being trained on relatively small datasets. In this paper, we present the workflow applied to prepare datasets of Romanian historical documents for training language-specific HTR models or for enhancing the language coverage of general purpose models. The process for creating such datasets uses the functionality implemented within the Transcribathon.eu tool for combining pre-existing manual transcriptions with automatically extracted HTR text. A preliminary evaluation of the effort required to create HTR training sets and an assessment of their quality is presented in this paper, allowing us to draw conclusions on the proposed approach.