Cannot make a desktop shortcut for the engine.exe file

[MatsLinder]

That is, I can make a shortcut but it does not work. Nothing happens when I double-click it. It would be very handy to have such a shortcut.

Since I first wrote this I discovered that if I place the shortcut on the task bar, it works. So it's all right (but still the matter of the desktop shortcut is strange).

And while I'm at it: where can I find explanations of the data for the zipped models on the Tatoeba MT Challenge page (i.e. what the actual values for BLEU, chr-F2 and test set size mean, for instance what are good values and not so good ones)?

[TommiNieminen]

The meaning of the BLEU and chr-F2 values is somewhat of an open problem, and unfortunately it depends a lot on the language pair. BLEU is the standard metric for MT, but it penalizes morphologically complex target languages (like Finnish) more than chr-F2, so the two metrics complement each other. Test set size is important because at very small sizes the results become unreliable (something like 1000 sentences should give fairly reliable results). One thing to note is that these scores are with the Tatoeba test set that is easier than most MT test sets, so usually in other publications you will see lower scores for the same language pairs.

Perhaps the best way to use the map is to select a target language, then see what score the model that translates from English to the selected target language achieves. This is usually the language direction with the most training resources. This score can be used as a benchmark to compare other language pairs. So for instance if you choose Swedish as target language, you can then compare the English source language BLEU of 62 to e.g. the Russian source language BLEU of 51.9, and infer that the Russian to Swedish model is not as good as the English to Swedish model, but probably still useful. On the other hand, the Japanese to Swedish model with a BLEU of 22.6 is so far below the English to Swedish model that it is probably not very good.

Btw. the Tatoeba models are not yet included in the download list in OPUS-CAT MT Engine, they will come in the next update (should be ready within a month). Currently you can download models from the OPUS-MT repository, which has a different mix of language pairs. However, the Tatoeba model scores for a language pair should be indicative of the OPUS-MT model quality for the same language pair, since the are trained with the same data with the same procedures.