LTR -- Lingvanex Test References for MT Evaluation from English into a total of 30 target languages for a big variety of cases.
| Parameter | Description |
|---|---|
| Length | Sentences from 1 to 100 words. |
| Domain | Medicine (12%), Automobile (11%), Finance (8%) |
| Tokenizer | Jupiter is 1.000.000 km far. Ask Mr. Johnson for training |
| Tags | I want to eat and swim |
| Capitalisation (Case) | HELLO my Dear frIEND |
| Different languages in one text (Up to 3 languages) | I see "Купалинка" performance near the theater. |
| Styling | Hello Dude! |
| Errors (Grammar, OCR) | I neet (need) to buy a kat (cat) |
| Abbreviations | The model was named 15.BVcX-10 |
| Named Entities | Let’s go to New York city |
| Idioms | A piece of cake. Once in a blue moon’ |
| Formulas (Math, Physics, Chemistry) | Cr2(SO4)3 + CO2 + H2O + K2SO4 + KNO3 |
| Romanian numbers | It was in MCMXVII year |
| Unicode Special Characters | №%&*/# |
LTR/Languages/Eng-* - folders with test references
LANGUAGES.tsv - mapping between ISO codes and language names
LICENSE.md
Data package has been released on 11/18/2023. Enjoy!
The LTR data set is released under the CC BY-SA 4.0 license.
@inproceedings{
title = "{LTR} Lingvanex Test References for {MT} Evaluation",
author = "Aliaksei Rudak",
url = "https://lingvanex.com",
Year = "2023"
}