Nyshporka 0.44: the new Pysar and Diak read confessional rolls and revision lists more accurately
Contents
Confessional rolls and revision lists are among the hardest books for a machine: dense columns of names, ages and households, written by hundreds of different hands. The new weights of both Cyrillic models were trained further on exactly such sheets, and now read them more accurately.
What has changed
Pysar v29 and Diak-Litopysets v12 were trained further on 41 sheets of confessional rolls from 1740–1910 and on 28 sheets of the revision of 1795. The confessional sheets were merged with the help of the FamilySearch name index: names and ages from the index helped to check each line.
Measurement was done on hands that were not in the training. The table shows the share of wrongly read characters; lower is better.
| What was read | Pysar v28 | Pysar v29 |
|---|---|---|
| Revision list of 1795 | 15.5 % | 12.5 % |
| Court book | 17.4 % | 15.7 % |
| Confessional rolls, four different hands | 17.0 % | 10.9 % |
| Parish registers and cases of the 19th century | 13.2 % | 13.2 % |
| Cases of the 20th century | 12.9 % | 12.8 % |
Surnames checked by eye in 2,808 records are found by the new weights just as well as by the previous ones. Diak-Litopysets, the second voice, makes fewer mistakes in the revision (17.3 → 16.2 %), and in confessional rolls 18.3 → 15.6 %.
Good to know
- Text read with the old weights remains valid. Re-reading already read cases with the new weights is not required.
- The old weights are not removed from the disk and keep reading until you install the new ones.
How to update
nysh update, then nysh migrate and nysh models get. Or a new nyshporka-setup.exe over the old one, and then nysh models get.
Or tell the assistant: "update Nyshporka and install the new weights".