Artificially generated triplets?

[unbabel]

Artificially generated triplets?

[captainvera]

Closing this since the issue has been solved.

Feel free to reopen if you have any further questions @MiroFurtado

[MiroFurtado]

oops - didn't mean to tag as bug!

[captainvera]

Hello @MiroFurtado

First of all, thanks for your interest in OpenKiwi!

You are right, the artificially generated training corpus (with ~500k triplets) we reference on our paper is not available publicly. It is based on the in-domain German corpus provided by WMT which is what we used to pre-train the predictors. We provide a link for this corpus in our Quickstart section here.

We had no plans for making this data available, but you raise a valid concern for reproducibility.
Thanks for bringing this to our attention!
We have to take some things into consideration but we will get back to you soon!

[MiroFurtado]

Great! Let me know what you end up deciding.

[trenous]

Actually the triplets are available in the data section on this site. The download there contains both the 500k triplets we used, and a larger corpus of 4 million triplets.

To generate QE tags from the triplets, you can follow the instructions in this repository

[trenous]

Hello @MiroFurtado ,

I uploaded the missing tags and sentence scores for the artificial roundtrip data to our releases.

If you wanted to recreate these files using the repository I posted in my previous comment, you need to train a FastAlign model, which is used to generate the source tags. We trained FastAlign on the English-German indomain corpus of 3 million parallel sentences, if you use a different FastAlign model you will get slightly different source tags.