Nyshporka 0.39: Bessarabian books in Romanian Cyrillic
Contents
Bessarabian books of the 1810s–1860s are often written in Romanian (Moldavian) Cyrillic. Until now Nyshporka read them poorly: the letters are familiar, but the language is foreign, so a record was hard to make out. This release adds a separate reader for such books, Logofet.
What kind of script this is
This is not a separate language but Romanian written in Cyrillic. In Russian documents it was usually called Moldavian. Romanian Cyrillic was used for centuries: the oldest securely dated Romanian document, the letter of Neacșu of Câmpulung of 1521, is written in it.
In 1860–1862 Wallachia and Moldavia switched to the Latin alphabet. Bessarabia had by then been part of the Russian Empire for half a century, so this reform did not touch it, and Cyrillic stayed here. Russian became the official language, and Romanian held on mainly in the church. After 1871 it was pushed out of there too.
So in the Orthodox books of Bessarabia, up to roughly the 1870s, you will come across this language: in parish registers and clergy returns, often next to Russian entries and signatures.
Who Logofet is
Logofet has learned one script: the Romanian Cyrillic of parish registers and clergy returns. It was trained on 17,870 lines from 196 Bessarabian cases.
Here is how often the readers make mistakes on records that volunteers read themselves, without Nyshporka's help:
| Reader | Errors per 100 letters | Roughly |
|---|---|---|
| Pysar | 45 | every second letter |
| Diak-Litopysets | 36 | every third |
| Logofet | 17 | every sixth |
This is 99 lines from seven sets. The sample is small, but the gap is large. Among them are 42 lines in two hands that Logofet has never seen: there it makes 12 errors per 100 letters, Diak-Litopysets 34, Pysar 39.
Another measurement concerns what you read a book for: names, surnames and villages. We counted how many of them were read so close to what is written that fuzzy search will find them:
| Records | Pysar | Logofet |
|---|---|---|
| difficult hands 1812–1839 | 35% | 58% |
| hands 1843–1865 | 74% | 85% |
These records were not part of the training. However, the reference for them was partly assembled with Logofet's help, so the first table is more reliable than the second.
What can be understood from such a text
We gave a human reader Logofet's output with no explanations: four pages of a clergy return of 1830. From the text alone it is clear what kind of document it is, which church and district, who is in the parish clergy, most of the names, wives and children, years of service. Ages and the tables of parishioners cannot be read without the scan.
Surnames remain the weakest point. The same surname in Romanian lines can fall apart, while in the Russian signature of the same clergy it is read almost correctly. That is why Logofet reads together with Pysar, not instead of it, and what is found is always worth checking against the frame.
Specialists
Logofet is the first in a new group of readers, specialists. A specialist is trained on one narrow script: on its own books it is better, and on all others it is worse. So it works differently from the ordinary readers:
- it does not download itself: not everyone needs it, and it takes up space;
- it does not switch itself on, even if it is already on your disk;
- you switch it on, for a specific book.
Nothing in your ordinary reading has changed. Whoever does not work with Bessarabia will not notice this change.
How to start
To see which specialists exist, which books they are for and which they are not for:
nysh models list
To download Logofet, once and about 60 MB:
nysh models get logofet-ro-v11
To read a book:
nysh read <справа> --with logofet_ro_v11.safetensors
Or just tell the assistant: "read this case with Logofet, it is in Romanian Cyrillic".
More details are on the page "Specialists".
How to update
nysh update, then nysh migrate. Or a new nyshporka-setup.exe over the old one.