Granular QE Tagging

[gsarti]

No description provided.

[gsarti]

Had a look at the progress and everything seems coherent to me! 👍 Some comments for final steps:

• I agree on the idea of adopting Levenshtein as fallback measure if embeddings for similarity are not provided.
• A new flag add_nametbd_quality_tags needs to be added to scripts/preprocess.py and parse_from_folder in divemt/parse_utils.py, following the approach used for WMT22 quality tags
• Since divemt/qe_taggers.py has become long and quite hard to navigate, it would make sense to create a divemt/taggers folder to separate its contents (a file for the QETagger ABC, a file for the WMT22QETagger and one for the NameTBDTagger, plus also tag_utils.py and wmt22qe_utils.py from the original folder). Paths in scripts will need to be adjusted accordingly!

The final goal is to be able to run preprocess.py to produce outputs containing QE annotations produced by the NameTBDTagger.

[k4black]

@gsarti Thanks for checking!

• Regarding Levenshtein I am more sure about hot to use it for mt-src
• It's currently full of debug code, that's why I do not commit it yet
• Sure! Make sense

[gsarti]

Yeah, Levenshtein can be used only for MT-PE matching, for MT-SRC embeddings should be required.