torchtext.data.metrics

bleu_score

torchtext.data.metrics.bleu_score(candidate_corpus, references_corpus, max_n=4, weights=[0.25, 0.25, 0.25, 0.25])

Computes the BLEU score between a candidate translation corpus and a references translation corpus. Based on https://www.aclweb.org/anthology/P02-1040.pdf

Parameters:

• candidate_corpus – an iterable of candidate translations. Each translation is an iterable of tokens
• references_corpus – an iterable of iterables of reference translations. Each translation is an iterable of tokens
• max_n – the maximum n-gram we want to use. E.g. if max_n=3, we will use unigrams, bigrams and trigrams
• weights – a list of weights used for each n-gram category (uniform by default)

Examples

>>> from torchtext.data.metrics import bleu_score
>>> candidate_corpus = [['I', 'ate', 'the', 'apple'], ['I', 'did']]
>>> references_corpus = [[['I', 'ate', 'it'], ['I', 'ate', 'apples']],
        [['I', 'did']]]
>>> bleu_score(candidate_corpus, references_corpus)
    0.7598356856515925