Small fix for SacreBLEUScore and the mteval-v13a tokenizer

[RistoAle97]

The behaviour of the mteval-v13a tokenizer (the ones used by WMT ) is the same as the original implementation now.

What does this PR do?

Fixes #1776

Before submitting

• Was this discussed/agreed via a Github issue? (no need for typos and docs improvements)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure to update the docs?
• Did you write any new necessary tests?

PR review

Anyone in the community is free to review the PR once the tests have passed.
If we didn't discuss your PR in Github issues there's a high chance it will not be merged.

Did you have fun?

Make sure you had fun coding 🙃

---

📚 Documentation preview 📚: https://torchmetrics--1778.org.readthedocs.build/en/1778/

[RistoAle97]

Some more insight on the effect of the change, I compared the hf version with the torchmetrics one (before and after applying my little change), here are the results on the newstest2013 en-es test set:

# translations is a sequence of generated spanish sentences, targets is a sequence of sequences of given spanish sentences
> hf_score = hf_scb.compute(predictions=translations, references=targets)
> hf_bleu_score = "{0:.5f}".format(hf_score["score"])
> print(f"HF implementation\nBLEU score: {hf_bleu_score}\npreds_len: {hf_score['sys_len']}\ntarget_len: {hf_score['ref_len']}")
... HF implementation
... BLEU score: 30.73243
... preds_len: 67090
... target_len: 70528

> scb_tm = SacreBLEUScore()  # current implementation
> tm_score = "{0:.5f}".format(scb_tm(translations, targets) * 100)
> print(f"Torchmetrics implementation\nBLEU score: {tm_score}\npreds_len: {scb_tm.preds_len}\ntarget_len: {scb_tm.target_len}")
... Torchmetrics implementation
... BLEU score: 30.67770
... preds_len: 67028.0
... target_len: 70468.0

> scb_tm_new = SacreBLEUScoreTest()  # my implementation with the applied change
> tm_score_new = "{0:.5f}".format(scb_tm_new(translations, targets) * 100)
> print(f"Torchmetrics implementation\nBLEU score: {tm_score_new}\npreds_len: {scb_tm_new.preds_len}\ntarget_len: {scb_tm_new.target_len}")
... Torchmetrics new version
... BLEU score: 30.73244
... preds_len: 67090.0
... target_len: 70528.0

I also compared the three versions on the WMT14 en-fr test set:

... Huggingface SacreBLEU: {'score': 33.17529924795069, 'counts': [48246, 29265, 19045, 12569], 'totals': [76495, 73492, 70489, 67487], 'precisions': [63.070788940453625, 39.82066075219071, 27.018400034047865, 18.624327648287817], 'bp': 0.9894540029829183, 'sys_len': 76495, 'ref_len': 77306}

... Torchmetrics SacreBLEU:
... score: 33.1378173828125
... preds_len: 76434.0
... target_len: 77247.0

... Torchmetrics SacreBLEU new version:
... score: 33.1753044128418
... preds_len: 76495.0
... target_len: 77306.0

while the BLEU score is still different, the discrepancy between the original version (used through HuggingFace) and the new torchmetrics implementation is smaller. The computed lengths for both predictions and references are the same now.

[codecov]

Codecov Report

Merging #1778 (50d97ba) into master (17c0e9f) will decrease coverage by 42%.
The diff coverage is 100%.

Additional details and impacted files

@@           Coverage Diff            @@
##           master   #1778     +/-   ##
========================================
- Coverage      87%     46%    -42%     
========================================
  Files         253     253             
  Lines       14164   14164             
========================================
- Hits        12387    6458   -5929     
- Misses       1777    7706   +5929