Detailed Information on Publication Record

Slovak Supermodel M1 (SSM1) On April 24, 2024 at 9:09 a.m., we launched the Transkribus Expert v. 1.27.0 creation of a new aggregate Slovak supermodel. The basis for the creation of a supermodel for certain Slovak manuscripts was the partial models of problem solvers in the Skriptor project The model has the designation ID63569 Slovak Supermodel M1 (SSM1). We used 2583 pages in Ground Truth (GT) quality to create the model; 56713 lines and 333777 words. Of these, 1224 pages for training (Train set) and 135 pages for verification of the new model (Validation set). The documents for this model are in Slovak, Latin, Hungarian and Czech, respectively. Slovakized Czech. The input manuscripts were of varying quality in terms of the digitized ones. Some digitized images were used from digital repositories, usually at a good quality of 600 dpi, and some were the result of scanning with the ScanTent device and DocScan software at a sufficient quality of 300 dpi. Samples of document fonts can be found in the Skriptor research study. We consider the demos themselves to be very important for the further use and improvement of the model, as other researchers will get an idea of the similar or different writing of their own documents that they want to transcribe. The values of Train set: 4.90% and Validation set: 5.30% were achieved. The creation of the SSM1 model on the Transkribus server took two days, 5 hours, 16 seconds, i.e. 53 hours and 58 minutes. The creation process was finished after 174 cycles (epochs). The SSM1 model is the first attempt in Slovakia and the Czech Republic to create a tool through which it would be possible to automatically make available certain types of handwritten documents that are similar to the fonts used to create them. In no case can the ID63569 Slovak Supermodel M1 (SSM1) be considered a definitive universal model for the transcription of historical manuscripts of Slovak provenance of all types and periods. The variety of fonts and handwriting styles is endless, and the creation of an optimal aggregated model is a challenge for other researchers and enthusiasts in the years to come. However, we believe that our first aggregated SSM1 model may facilitate the automatic transcription of further analogous manuscripts. The research team plans to make the datasets available as part of the sustainability of the project in 2024-2028, preferably for research and educational purposes for institutions and researchers who want to contribute to the creation of a model of historical manuscripts in West Slavic languages, respectively languages of Slovak and Bohemian origin. Copyright: CC BY-NC-SA. Of course, such automatic transcription will not immediately bring satisfactory results. However, it can facilitate the "rough" step-by-step automatic transcription of additional pages, manually correcting them to GT status, and then using larger GT datasets to refine a new model based on our SSM1. After the creation of hundreds and thousands more pages of GT, it will be possible to approach the creation of next generations of new models based on SSM1. Development could continue for manuscripts with models of new generations SSM2, SSM3 etc. Respectively, for prints and typesets (P-Print) such as supermodels SSP1, SSP2, SSP3 etc. (Google translate)