Чи имля

Чи имля

Имлядә язылган татарча басма һәм кулъязмаларны уку семинары.
SkyTat мәктәбе ярдәмендә үтә.

imlaTAT OCR project news

imlaTATriza

imlaTAT 0.7 fine-tuned on 156 lines of Rizaeddin Fakhreddin’s handwritten Tärjemäm. On a held-out page the character error is 21 % (0.7 had 36 %) and the word error is 62 % (0.7 had 89 %).

imlaTAT 0.7

Trained on 3k proofread lines from seven printed editions of 1906 to 1935. On held-out lines the character error is 3.4 % and the word error is 17 %. With the lexicon decoder they are 2.9 % and 12.6 %. Reads Arabic-script Old Tatar and outputs Cyrillic. Runs offline, 0.1 to 0.2 s per line.

imlaTAT 0.6

OsmanLat (Ottoman to Latin Turkish model) retrained on 0,7k lines from two editions to output Cyrillic. On held-out lines the character error is 5 % and the word error is 27 %. Untrained OsmanLat on the same lines gives 91 % character error.