corpus [prepared by Pierre Lison and Jörg Tiedemann](http://stp.lingfil.uu.se/~joerg/paper/opensubs2016.pdf).
Metric
------
[BLEU](https://en.wikipedia.org/wiki/BLEU) is used as the evaluation metric.
Directory structure
-------------------
*`README.md` — this file
*`config.txt` — GEval configuration file
*`train/` — directory with training data
*`train/train.tsv` — train set (Russian-Polish corpus of 2M sentence pairs, the Polish sentence is given in the first column, the Russian sentence — in the second one)
*`dev-0/` — directory with development data (20K sentence pairs)
*`dev-0/in.tsv` — input data for the dev set (Russian utterances)
*`dev-0/expected.tsv` — expected (reference) data for the dev set (Polish utterances)
*`test-A` — directory with test data (20K sentence pairs)
*`test-A/in.tsv` — input data for the test set
*`test-A/expected.tsv` — expected data for the test set (not available in the master branch)
The Polish utterances in `{dev-0,test-A}/expected.tsv` are tokenized
(by [Moses](http://www.statmt.org/moses/) tokenizer.perl script, with
`-no-escape-option`), but the Russian utterances in all the files and
Polish utterances in `train/train.tsv` are *NOT* tokenized.
Note that some of the files have `.tsv` extension, even though they
actually do not contain TABs — this is just for the compatibility with
