Use right range and threshold for showing "bad" words/sentences

[abhi-agg]

Higher QE scores means better quality. Changed the threshold to -0.5 ln(0.5) => -0.6931 as per discussions in QE meetings.

@mfomicheva @abarbosa94 @felipesantosk Please let me know if any of the above is wrong 👍🏾

[jelmervdl]

If these are the correct thresholds we can also lose the "these thresholds are just examples" comments

[abhi-agg]

I just want to confirm with @mfomicheva @abarbosa94 @felipesantosk once more that the threshold of -0.5 is a good one as a starting point for all the language pairs irrespective of whether the quality scores are returned using translation models or supervised QE models under the hood.

I can remove the comment after their confirmation. Thanks for pointing out 👍🏾

[mfomicheva]

I just want to confirm with @mfomicheva @abarbosa94 @felipesantosk once more that the threshold of -0.5 is a good one as a starting point for all the language pairs irrespective of whether the quality scores are returned using translation models or supervised QE models under the hood.

I can remove the comment after their confirmation. Thanks for pointing out 👍🏾

I responded on slack

[abhi-agg]

Just documenting what @mfomicheva shared:

For the supervised models that were fitted on annotated data (En-Es, En-Cs and En-Et language pairs), you should use the threshold that corresponds to the log of 0.5, which is around -0.6931 (here log means ln)

For the unsupervised case where the returned value is just the average log-prob coming directly from the MT model, I think you should still start with the same threshold and experiment further with it

The range [-0.6931, 0] means better quality

I will modify PR to reflect these changes. Updating the description of the PR as well.