[Doc] explaining romanian postprocessing for MBART BLEU hacking

[sshleifer]

No description provided.

[sgugger]

LGTM

[codecov]

Codecov Report

Merging #5943 into master will decrease coverage by 0.00%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #5943      +/-   ##
==========================================
- Coverage   77.31%   77.31%   -0.01%     
==========================================
  Files         146      146              
  Lines       26214    26214              
==========================================
- Hits        20268    20267       -1     
- Misses       5946     5947       +1     

Impacted Files	Coverage Δ
src/transformers/file_utils.py	81.19% <0.00%> (-0.30%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update ccbf74a...163ba92. Read the comment docs.

[vince62s]

@sshleifer just a quick comment on this, from here: facebookresearch/fairseq#1758 (comment)

for EN to RO, there is no point of maximising BLEU by removing diacritics when in real world (WMT human evaluation + SMT Matrix) it is clearly compared with a reference which HAS diacritics. But lot of papers do not do the things right.

[sshleifer]

That makes sense @vince62s! This is mostly so I can have a link to paste into a github issue when people ask me why their BLEU score is 27 :)