Skip to content

Latest commit

 

History

History
88 lines (72 loc) · 3.06 KB

README.md

File metadata and controls

88 lines (72 loc) · 3.06 KB

Build Status

language-detector

language-detector detects the language of text

Installation

pip install language-detector

Python Version

Works with both Python 2 and 3

Use

from language_detector import detect_language
text = "I arrived in that city on January 4, 1937"
language = detect_language(text)
# prints English

Features

Languages Supported
Arabic
English
Farsi
French
German
Khmer
Kurmanci (Kurdish)
Mandarin
Russian
Sorani (Kurdish)
Spanish
Turkish

Testing

To test the package run

python -m unittest language_detector.tests.test

Comparison

Test is a comparison of how well language-detector and langid identify languages in the data sources.

package language-detector langid
test-duration (in seconds) 0.10 3.83
accuracy 96.77% 67.74%

Excluding Languages

If you don't want language-detector to look for certain languages, you can monkey-patch the code. For example, in order to exclude English:

import language_detector
language_detector.char_language = [cl for cl in char_language if cl[1] != "English"]

# proceed as normal

Datasets

The following is a list of datasets used for each language:

Language Datasets
Arabic UN Corpora
English UN Corpora
Farsi BBC News Persian
French UN Corpora
German Deutsche Welle
Khmer Cambodia Daily
Kurmanci (Kurdish) Rudaw
Mandarin UN Corpora
Russian UN Corpora
Sorani (Kurdish) Rudaw
Spanish UN Corpora
Turkish BBC News Türkçe

How Does It Work?

When training the model, we scan all the data sources and compute the frequency of how often a character appears in each specific language. We also compute the frequency of how often a characters appears in all of the data sources for all the languages. For each language, we then calculate a score for each character as frequency_in_language / frequency_in_all_languages. We then save the top ten highest scoring characters for each language.
When detecting a language, we simply iterate through the saved characters (ten for each language), and add their score as a weighted-vote for each language. Whichever, language has the highest score is selected as the winner.

Contributing

If you'd like to contribute a new language, please consult CONTRIBUTING.md

Support

Contact the package author, Daniel J. Dufour, at daniel.j.dufour@gmail.com