Implementation for paper:

BLEU: a Method for Automatic Evaluation of Machine Translation

Author: Ba Ngoc from ProtonX

BLEU score is a popular metric to evaluate machine translation. Check out the recent Transformer project we published.

I. Usage

from bleu_score import cal_corpus_bleu_score

candidates = ['eating chicken chicken is a eating a eating chicken',
              'eating chicken chicken is not good']
references_list = [['a chicken is eating chicken', 'there is a chicken eating chicken'], [
    'a chicken is eating chicken', 'there is a chicken eating chicken']]

bleu_score = cal_corpus_bleu_score(candidates, references_list,
                      weights=(0.25, 0.25, 0.25, 0.25), N=4)

print('Bleu Score: {}'.format(bleu_score))

II. BLEU Score Formula

We count specific n-grams in the candidates and the number of those grams in the references. Then we calculate the proportion of two countings and get the precision.

Important to note: Count clip means that the number of typical n-grams can not exceed the maximum number of that n-grams in any single reference.

For example: if ('a', 'a') gram exists 3 times in a candidate. However, the maximum number of this gram in any single reference is 2. So we will use value 2 for calculation.

If you never heard about grams? It means that we count the number of continuous substrings with a pre-set length in a string.

Candidate 1: 'eating chicken chicken is a eating a eating chicken'

-------Unigram------

-------bigrams------

We can do the same thing with trigrams and 4-grams

We prefer the reference with a length that is closest to the candidate's.

Checkout function get_eff_ref_length in utils.py.

c: the total lengths of all candidates

r: the total lengths of all effective reference lengths

N: the number of grams

w: list of pre-set weight for each gram