pytorch / text

Models, data loaders and abstractions for language processing, powered by PyTorch
https://pytorch.org/text
BSD 3-Clause "New" or "Revised" License
3.48k stars 816 forks source link

BLEU_SCORE weird behaviour #2224

Open qlemesle opened 5 months ago

qlemesle commented 5 months ago

❓ Questions and Help

Is it normal that with standard bleu settings if an n_gram isn't in references then the score returned is 0 ?

With default settings (4 n_grams and weights = [0.025] *4) The candidate : ['Le', 'petit', 'chat', 'moche'] For the reference : ['Le', 'petit', 'chat', 'mignon'] Get a bleu score of 0.

Due to this line in the code : if min(clipped_counts) == 0: return 0.0

Is this behaviour normal ?