Skip to main content

A Python Wrapper to calculate standard BLEU scores for NLP

Project description

bleu (Python Package)

Pypi Downloads Month_Downloads

A Python Wrapper for the standard BLEU evaluation for Natural Language Generation (NLG).


Requirement: Python 3

Option 1: Install pip package

pip install --upgrade bleu

Option 2: Build from source

pip install --upgrade git+git://

How to Run

The most standard way to calculate BLEU is by Moses' script for detokenized BLEU. This package provides easy calls to it.

Function 1: Calculate the BLEU for lists

If you want to check only one hypothesis (a list of sentences):

>>> from bleu import list_bleu
>>> ref = ['it is a white cat .',
             'wow , this dog is huge .']
>>> ref1 = ['This cat is white .',
             'wow , this is a huge dog .']
>>> hyp = ['it is a white kitten .',
            'wowww , the dog is huge !']
>>> hyp1 = ["it 's a white kitten .",
             'wow , this dog is huge !']
>>> list_bleu([ref], hyp)
>>> list_bleu([ref, ref1], hyp1)

If you want to check multiple hypothesis (several lists of sentences):

>>> from bleu import multi_list_bleu
>>> multi_list_bleu([ref, ref1], [hyp, hyp1])
[34.99, 57.91]

detok=False: It is not advisable to use tokenized bleu (by multi-bleu.perl), but if you want to call it, just use detok=False:

>>> list_bleu([ref], hyp, detok=False)
# or if you want to test multiple hypotheses
>>> multi_list_bleu([ref, ref1], [hyp, hyp1], detok=False)
[39.76, 47.47]

verbose=True: If there are unexpected errors, you might want to check the intermediate steps by verbose=True.

Function 2: Calculate the BLEU for files

If you want to check only one hypothesis file:

# if you already have the following files
>>> from bleu import file_bleu
>>> hyp_file = 'data/hyp0.txt'
>>> ref_files = ['data/ref0.txt', 'data/ref1.txt']
>>> file_bleu(ref_files, hyp_file)

If you want to check multiple hypothesis files:

>>> from bleu import multi_file_bleu
>>> hyp_file1 = 'data/hyp1.txt'
>>> bleus = multi_file_bleu(ref_files, [hyp_file, hyp_file1])
[34.99, 57.91]

detok=True: Set it if you want to calculate the (not recommended) tokenized bleu.

verbose=True: Set it if you want to inspect how the bleu calculations are made:

>>> bleu = file_bleu(ref_files, hyp_file, verbose=True)
[Info] Valid Reference Files: ['data/ref0.txt', 'data/ref1.txt']
[Info] Valid Hypothesis Files: ['data/hyp0.txt']
[Info] #lines in each file: 2
[cmd] perl detokenizer.perl -l en < data/ref0.txt > data/ref0.detok.txt 2>/dev/null
[cmd] perl detokenizer.perl -l en < data/ref1.txt > data/ref1.detok.txt 2>/dev/null
[cmd] perl detokenizer.perl -l en < data/hyp0.txt > data/hyp0.detok.txt 2>/dev/null
[cmd] perl multi-bleu-detok.perl data/ref0.detok.txt data/ref1.detok.txt < data/hyp0.detok.txt
2-ref bleu for data/hyp0.detok.txt: 34.99
>>> bleu

Option 3: Detokenize files

>>> from bleu import detok_files
>>> detok_ref_files = detok_files(ref_files, tmp_dir='./data', file_prefix='ref_dtk', verbose=True)
[cmd] perl ./TMP_DIR/detokenizer.perl -l en < data/ref0.txt > data/ref_dtk0.txt 2>/dev/null
[cmd] perl ./TMP_DIR/detokenizer.perl -l en < data/ref1.txt > data/ref_dtk1.txt 2>/dev/null
>>> detok_ref_files
['data/ref_dtk0.txt', 'data/ref_dtk1.txt']

In Case of Unexpected Outputs

Check the python file and adapt it.


If you have more questions, feel free to check out the common Q&A, or raise a new GitHub issue.

In case of really urgent needs, contact the author Zhijing Jin (Miss).

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for bleu, version 0.3
Filename, size File type Python version Upload date Hashes
Filename, size bleu-0.3.tar.gz (5.2 kB) File type Source Python version None Upload date Hashes View

Supported by

Pingdom Pingdom Monitoring Google Google Object Storage and Download Analytics Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page