TraPrInq Project: May 2023, eighth training of the HTR Model
The goal of the TraPrInq project is to create a HTR generic model that will allow the automatic transcription of the documents from the archives (court trials, “processos”) of the Portuguese Inquisition. Since mid-January 2022, our team of paleographers has transcribed more than one million words from a very large number of hands and all periods from the early 16th century to the early nineteenth century.
On the previous training and model (M-7), see here.
In terms of CER (Character Error Rate) overall in the Validation set (VS, 493 untrained pages, totalizing 105,339 words) and number of trained pages in the training set (TS, 5269 pages, 1,076 370 words), the result of the new model trained in the engine Pylaia (M-8.2) is 6.1%.
Some remarks on Pylaia Model trained in May.
The system generates 2 CERs:
- a) the one that comes in the model description: 1%;
- b) what comes in line 1 (Overall) of the xls sheet of the VS results: 6,61%.
- c) Furthermore, a third CER can be arithmetically obtained by averaging the total CERs per page and dividing the total average by the number of pages in the VS (now 493 pages): 6.75%.
Contrary to the previous results, all illustrate a slight decrease of character accuracy (or conversely an increase of the CER) compared to the previous models M-6 (a: 5.5%; b: 6.01%) and M-7 (a: 5.5%; b: 5.87%).
According to the third calculation method (c), 8 pages show a CER superior to 20%; no one improved with the new trained model and they are responsible for half a point of the CER overall (6,28%).
28 pages (5.6% of the total of the VS) with a CER higher than 15% (from 15.29 to 78.8%) never improved along the trainings. They are responsible for one point of the CER overall (5,80%).
226 pages (46% of the VS) with an individual CER inferior to 5% show an overall of 3,5%.
321 pages (65% of the VS) with an individual CER inferior to 10% show an overall of 4,40%.
Conclusions
Compared to the previous trained models, there has been a slight slowdown in the CER results. Some pages with near results since M-5 strongly increased their CER after this training, for example, p. 19 (6%/10%), 240 (9%/16%), 418 (40/48%).
OpenEdition vous propose de citer ce billet de la manière suivante :
Hervé Baudry (4 juin 2023). TraPrInq Project: May 2023, eighth training of the HTR Model. e-Inquisition. Consulté le 24 mai 2025 à l’adresse https://doi.org/10.58079/uubo