NLTK: corpus-level bleu vs sentence-level BLEU score

Question

I have imported nltk in python to calculate BLEU Score on Ubuntu. I understand how sentence-level BLEU score works, but I don't understand how corpus-level BLEU score work.

Below is my code for corpus-level BLEU score:

import nltk

hypothesis = ['This', 'is', 'cat'] 
reference = ['This', 'is', 'a', 'cat']
BLEUscore = nltk.translate.bleu_score.corpus_bleu([reference], [hypothesis], weights = [1])
print(BLEUscore)

For some reason, the bleu score is 0 for the above code. I was expecting a corpus-level BLEU score of at least 0.5.

Here is my code for sentence-level BLEU score

import nltk

hypothesis = ['This', 'is', 'cat'] 
reference = ['This', 'is', 'a', 'cat']
BLEUscore = nltk.translate.bleu_score.sentence_bleu([reference], hypothesis, weights = [1])
print(BLEUscore)

Here the sentence-level BLEU score is 0.71 which I expect, taking into account the brevity-penalty and the missing word "a". However, I don't understand how corpus-level BLEU score work.

Any help would be appreciated.

Answer 1

TL;DR:

>>> import nltk
>>> hypothesis = ['This', 'is', 'cat'] 
>>> reference = ['This', 'is', 'a', 'cat']
>>> references = [reference] # list of references for 1 sentence.
>>> list_of_references = [references] # list of references for all sentences in corpus.
>>> list_of_hypotheses = [hypothesis] # list of hypotheses that corresponds to list of references.
>>> nltk.translate.bleu_score.corpus_bleu(list_of_references, list_of_hypotheses)
0.6025286104785453
>>> nltk.translate.bleu_score.sentence_bleu(references, hypothesis)
0.6025286104785453

(Note: You have to pull the latest version of NLTK on the develop branch in order to get a stable version of the BLEU score implementation)

---

In Long:

Actually, if there's only one reference and one hypothesis in your whole corpus, both corpus_bleu() and sentence_bleu() should return the same value as shown in the example above.

In the code, we see that sentence_bleu is actually a duck-type of corpus_bleu:

def sentence_bleu(references, hypothesis, weights=(0.25, 0.25, 0.25, 0.25),
                  smoothing_function=None):
    return corpus_bleu([references], [hypothesis], weights, smoothing_function)

And if we look at the parameters for sentence_bleu:

 def sentence_bleu(references, hypothesis, weights=(0.25, 0.25, 0.25, 0.25),
                      smoothing_function=None):
    """"
    :param references: reference sentences
    :type references: list(list(str))
    :param hypothesis: a hypothesis sentence
    :type hypothesis: list(str)
    :param weights: weights for unigrams, bigrams, trigrams and so on
    :type weights: list(float)
    :return: The sentence-level BLEU score.
    :rtype: float
    """

The input for sentence_bleu's references is a list(list(str)).

So if you have a sentence string, e.g. "This is a cat", you have to tokenized it to get a list of strings, ["This", "is", "a", "cat"] and since it allows for multiple references, it has to be a list of list of string, e.g. if you have a second reference, "This is a feline", your input to sentence_bleu() would be:

references = [ ["This", "is", "a", "cat"], ["This", "is", "a", "feline"] ]
hypothesis = ["This", "is", "cat"]
sentence_bleu(references, hypothesis)

When it comes to corpus_bleu() list_of_references parameter, it's basically a list of whatever the sentence_bleu() takes as references:

def corpus_bleu(list_of_references, hypotheses, weights=(0.25, 0.25, 0.25, 0.25),
                smoothing_function=None):
    """
    :param references: a corpus of lists of reference sentences, w.r.t. hypotheses
    :type references: list(list(list(str)))
    :param hypotheses: a list of hypothesis sentences
    :type hypotheses: list(list(str))
    :param weights: weights for unigrams, bigrams, trigrams and so on
    :type weights: list(float)
    :return: The corpus-level BLEU score.
    :rtype: float
    """

Other than look at the doctest within the nltk/translate/bleu_score.py, you can also take a look at the unittest at nltk/test/unit/translate/test_bleu_score.py to see how to use each of the component within the bleu_score.py.

By the way, since the sentence_bleu is imported as bleu in the (nltk.translate.__init__.py](https://github.com/nltk/nltk/blob/develop/nltk/translate/init.py#L21), using

from nltk.translate import bleu 

would be the same as:

from nltk.translate.bleu_score import sentence_bleu

and in code:

>>> from nltk.translate import bleu
>>> from nltk.translate.bleu_score import sentence_bleu
>>> from nltk.translate.bleu_score import corpus_bleu
>>> bleu == sentence_bleu
True
>>> bleu == corpus_bleu
False

Answer 2

Let's take a look:

>>> help(nltk.translate.bleu_score.corpus_bleu)
Help on function corpus_bleu in module nltk.translate.bleu_score:

corpus_bleu(list_of_references, hypotheses, weights=(0.25, 0.25, 0.25, 0.25), smoothing_function=None)
    Calculate a single corpus-level BLEU score (aka. system-level BLEU) for all 
    the hypotheses and their respective references.  

    Instead of averaging the sentence level BLEU scores (i.e. marco-average 
    precision), the original BLEU metric (Papineni et al. 2002) accounts for 
    the micro-average precision (i.e. summing the numerators and denominators
    for each hypothesis-reference(s) pairs before the division).
    ...

You're in a better position than me to understand the description of the algorithm, so I won't try to "explain" it to you. If the docstring does not clear things up enough, take a look at the source itself. Or find it locally:

>>> nltk.translate.bleu_score.__file__
'.../lib/python3.4/site-packages/nltk/translate/bleu_score.py'