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.
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.classifyis 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.
- Install the package:
pip3 install py3langid(orpipwhere applicable)
- 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')# 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)The docs below are provided for reference, only part of the functions are currently tested and maintained.
langid.py is a standalone Language Identification (LangID) tool.
The design principles are as follows:
- Fast
- Pre-trained over a large number of languages (currently 97)
- Not sensitive to domain-specific features (e.g. HTML/XML markup)
- Single .py file with minimal dependencies
- 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
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)
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.
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)
So far Python 2.7 only, see the original instructions.
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