py3langid 0.3.0

Fork of the language identification tool langid.py, featuring a modernized codebase and faster execution times.

py3langid is a fork of the standalone language identification tool langid.py by Marco Lui.

Original license: BSD-2-Clause. Fork license: BSD-3-Clause.

Changes in this fork

Execution speed has been improved and the code base has been optimized for Python 3.6+:

• Import: Loading the package (import py3langid) is about 30% faster

• Startup: Loading the default classification model is 25-30x faster

• Execution: Language detection with langid.classify is 5-6x faster on paragraphs (less on longer texts)

For implementation details see this blog post: How to make language detection with langid.py faster.

For more information and older Python versions see changelog.

Usage

Drop-in replacement

1. Install the package:

   • pip3 install py3langid (or pip where applicable)

2. Use it:

   • with Python: import py3langid as langid

   • on the command-line: langid

With Python

Basics:

>>> import py3langid as langid

>>> text = 'This text is in English.'
# identified language and probability
>>> langid.classify(text)
('en', -56.77429)
# unpack the result tuple in variables
>>> lang, prob = langid.classify(text)
# all potential languages
>>> langid.rank(text)

More options:

>>> from py3langid.langid import LanguageIdentifier, MODEL_FILE

# subset of target languages
>>> identifier = LanguageIdentifier.from_pickled_model(MODEL_FILE)
>>> identifier.set_languages(['de', 'en', 'fr'])
# this won't work well...
>>> identifier.classify('这样不好')
('en', -81.831665)

# normalization of probabilities to an interval between 0 and 1
>>> identifier = LanguageIdentifier.from_pickled_model(MODEL_FILE, norm_probs=True)
>>> identifier.classify('This should be enough text.')
('en', 1.0)

Note: the Numpy data type for the feature vector has been changed to optimize for speed. If results are inconsistent, try restoring the original setting:

>>> langid.classify(text, datatype='uint32')

On the command-line

# basic usage with probability normalization
$ echo "This should be enough text." | langid -n
('en', 1.0)

# define a subset of target languages
$ echo "This won't be recognized properly." | langid -n -l fr,it,tr
('it', 0.97038305)

Legacy documentation

The docs below are provided for reference, only part of the functions are currently tested and maintained.

Introduction

langid.py is a standalone Language Identification (LangID) tool.

The design principles are as follows:

1. Fast

2. Pre-trained over a large number of languages (currently 97)

3. Not sensitive to domain-specific features (e.g. HTML/XML markup)

4. Single .py file with minimal dependencies

5. Deployable as a web service

All that is required to run langid.py is Python >= 3.6 and numpy.

The accompanying training tools are still Python2-only.

langid.py is WSGI-compliant. langid.py will use fapws3 as a web server if available, and default to wsgiref.simple_server otherwise.

langid.py comes pre-trained on 97 languages (ISO 639-1 codes given):

af, am, an, ar, as, az, be, bg, bn, br, bs, ca, cs, cy, da, de, dz, el, en, eo, es, et, eu, fa, fi, fo, fr, ga, gl, gu, he, hi, hr, ht, hu, hy, id, is, it, ja, jv, ka, kk, km, kn, ko, ku, ky, la, lb, lo, lt, lv, mg, mk, ml, mn, mr, ms, mt, nb, ne, nl, nn, no, oc, or, pa, pl, ps, pt, qu, ro, ru, rw, se, si, sk, sl, sq, sr, sv, sw, ta, te, th, tl, tr, ug, uk, ur, vi, vo, wa, xh, zh, zu

The training data was drawn from 5 different sources:

• JRC-Acquis

• ClueWeb 09

• Wikipedia

• Reuters RCV2

• Debian i18n

Usage

langid [options]

optional arguments:

-h, --help

show this help message and exit

-s, --serve

launch web service

--host=HOST

host/ip to bind to

--port=PORT

port to listen on

-v

increase verbosity (repeat for greater effect)

-m MODEL

load model from file

-l LANGS, --langs=LANGS

comma-separated set of target ISO639 language codes (e.g en,de)

-r, --remote

auto-detect IP address for remote access

-b, --batch

specify a list of files on the command line

-d, --dist

show full distribution over languages

-u URL, --url=URL

langid of URL

--line

process pipes line-by-line rather than as a document

-n, --normalize

normalize confidence scores to probability values

The simplest way to use langid.py is as a command-line tool, and you can invoke using python langid.py. If you installed langid.py as a Python module (e.g. via pip install langid), you can invoke langid instead of python langid.py -n (the two are equivalent). This will cause a prompt to display. Enter text to identify, and hit enter:

>>> This is a test
('en', -54.41310358047485)
>>> Questa e una prova
('it', -35.41771221160889)

langid.py can also detect when the input is redirected (only tested under Linux), and in this case will process until EOF rather than until newline like in interactive mode:

python langid.py < README.rst
('en', -22552.496054649353)

The value returned is the unnormalized probability estimate for the language. Calculating the exact probability estimate is disabled by default, but can be enabled through a flag:

python langid.py -n < README.rst
('en', 1.0)

More details are provided in this README in the section on Probability Normalization.

You can also use langid.py as a Python library:

# python
Python 2.7.2+ (default, Oct  4 2011, 20:06:09)
[GCC 4.6.1] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import langid
>>> langid.classify("This is a test")
('en', -54.41310358047485)

Finally, langid.py can use Python’s built-in wsgiref.simple_server (or fapws3 if available) to provide language identification as a web service. To do this, launch python langid.py -s, and access http://localhost:9008/detect . The web service supports GET, POST and PUT. If GET is performed with no data, a simple HTML forms interface is displayed.

The response is generated in JSON, here is an example:

{"responseData": {"confidence": -54.41310358047485, "language": "en"}, "responseDetails": null, "responseStatus": 200}

A utility such as curl can be used to access the web service:

# curl -d "q=This is a test" localhost:9008/detect
{"responseData": {"confidence": -54.41310358047485, "language": "en"}, "responseDetails": null, "responseStatus": 200}

You can also use HTTP PUT:

# curl -T readme.rst localhost:9008/detect
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                               Dload  Upload   Total   Spent    Left  Speed
100  2871  100   119  100  2752    117   2723  0:00:01  0:00:01 --:--:--  2727
{"responseData": {"confidence": -22552.496054649353, "language": "en"}, "responseDetails": null, "responseStatus": 200}

If no “q=XXX” key-value pair is present in the HTTP POST payload, langid.py will interpret the entire file as a single query. This allows for redirection via curl:

# echo "This is a test" | curl -d @- localhost:9008/detect
{"responseData": {"confidence": -54.41310358047485, "language": "en"}, "responseDetails": null, "responseStatus": 200}

langid.py will attempt to discover the host IP address automatically. Often, this is set to localhost(127.0.1.1), even though the machine has a different external IP address. langid.py can attempt to automatically discover the external IP address. To enable this functionality, start langid.py with the -r flag.

langid.py supports constraining of the output language set using the -l flag and a comma-separated list of ISO639-1 language codes (the -n flag enables probability normalization):

# python langid.py -n -l it,fr
>>> Io non parlo italiano
('it', 0.99999999988965627)
>>> Je ne parle pas français
('fr', 1.0)
>>> I don't speak english
('it', 0.92210605672341062)

When using langid.py as a library, the set_languages method can be used to constrain the language set:

python
Python 2.7.2+ (default, Oct  4 2011, 20:06:09)
[GCC 4.6.1] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import langid
>>> langid.classify("I do not speak english")
('en', 0.57133487679900674)
>>> langid.set_languages(['de','fr','it'])
>>> langid.classify("I do not speak english")
('it', 0.99999835791478453)
>>> langid.set_languages(['en','it'])
>>> langid.classify("I do not speak english")
('en', 0.99176190378750373)

Batch Mode

langid.py supports batch mode processing, which can be invoked with the -b flag. In this mode, langid.py reads a list of paths to files to classify as arguments. If no arguments are supplied, langid.py reads the list of paths from stdin, this is useful for using langid.py with UNIX utilities such as find.

In batch mode, langid.py uses multiprocessing to invoke multiple instances of the classifier, utilizing all available CPUs to classify documents in parallel.

Probability Normalization

The probabilistic model implemented by langid.py involves the multiplication of a large number of probabilities. For computational reasons, the actual calculations are implemented in the log-probability space (a common numerical technique for dealing with vanishingly small probabilities). One side-effect of this is that it is not necessary to compute a full probability in order to determine the most probable language in a set of candidate languages. However, users sometimes find it helpful to have a “confidence” score for the probability prediction. Thus, langid.py implements a re-normalization that produces an output in the 0-1 range.

langid.py disables probability normalization by default. For command-line usages of langid.py, it can be enabled by passing the -n flag. For probability normalization in library use, the user must instantiate their own LanguageIdentifier. An example of such usage is as follows:

>> from py3langid.langid import LanguageIdentifier, MODEL_FILE
>> identifier = LanguageIdentifier.from_pickled_model(MODEL_FILE, norm_probs=True)
>> identifier.classify("This is a test")
('en', 0.9999999909903544)

Training a model

So far Python 2.7 only, see the original instructions.

Read more

langid.py is based on published research. [1] describes the LD feature selection technique in detail, and [2] provides more detail about the module langid.py itself.

[1] Lui, Marco and Timothy Baldwin (2011) Cross-domain Feature Selection for Language Identification, In Proceedings of the Fifth International Joint Conference on Natural Language Processing (IJCNLP 2011), Chiang Mai, Thailand, pp. 553—561. Available from http://www.aclweb.org/anthology/I11-1062

[2] Lui, Marco and Timothy Baldwin (2012) langid.py: An Off-the-shelf Language Identification Tool, In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (ACL 2012), Demo Session, Jeju, Republic of Korea. Available from www.aclweb.org/anthology/P12-3005