Language Detection Library - 99% over precision for 49 languages - 12/3/2010 Nakatani Shuyo @ Cybozu Labs, Inc.
What languages are these? sprogregistrering
What languages are these? sprogregistrering Danish
What languages are these? ةغللا نع فشكلا
تخانش یک نابز
What languages are these? Arabic ةغللا نع فشكلا
Persian نابز صیخشت
تخانش یک نابز Urdu
What languages are these? Language Detection
What languages are these? Language Detection English
What’s “Language Detection”? Detect language in which texts are written
also character code detection (excluded) alias: Language identification / Language guessing
Japanese English Chinese German Spanish Italian Arabic Hindi Korean
Why Language Detection? Purpose
For language of search criteria Query “Java” => Hit Chinese texts...
For SPAM filter/Extract content filter To use language-specific information(punctuations, keywords) Usage
Web search engine Apache Nutch bundles a language detection module
Bulletin board Post in English, Japanese and Chinese
Methods The more languages, the more difficult
Among languages with the same script Requires knowledge of scripts and languages A simple method:
Matching with the dictionary in each language Huge dictionary(inflections, compound words)
Calcurates language probabilities from features of spelling Naive Bayse with character n-gram
Existing Language Detection There are a few libraries of language
detection. Usage was limited? For only web search?
But all services will become global from now on!
Building corpus/model is a expensive work. Requires knowledge of scripts and languages
Few languages supported & low precision
Almost 10 languages. Not including Asian ones
“Practical” Language Detection 99% over precision
90% is not practical. (100 of 1000 mistakes) 50 languages supported
European, Asian and so on Fast Detection
Many documents available Output each language’s probability
For multiple candidates
Language Detection Library for Java We developed a language detection library for Java. Generates the language profiles from training corpus Profile : the probabilities of all spellings in each language
Returns the candidates and their probabilities for given texts 49 languages supported Open Source (Apache License 2.0)
49 languages from Wikipedia That can provide a test corpus of its language
200 news articles of 49 languages Google News (24 languages)
News sites in each language
Crawling by RSS
languages # precisions items af Afrikaans 200 199 (99.50%) en=1, af=199 ar Arabic 200 200 (100.00%) ar=200 bg Bulgarian 200 200 (100.00%) bg=200 bn Bengali 200 200 (100.00%) bn=200 cs Czech 200 200 (100.00%) cs=200 da Danish 200 179 (89.50%) da=179, no=14, en=7 de German 200 200 (100.00%) de=200 el Greek 200 200 (100.00%) el=200 en English 200 200 (100.00%) en=200 es Spanish 200 200 (100.00%) es=200 fa Persian 200 200 (100.00%) fa=200 fi Finnish 200 200 (100.00%) fi=200 fr French 200 200 (100.00%) fr=200 gu Gujarati 200 200 (100.00%) gu=200 he Hebrew 200 200 (100.00%) he=200 hi Hindi 200 200 (100.00%) hi=200 hr Croatian 200 200 (100.00%) hr=200 hu Hungarian 200 200 (100.00%) hu=200 id Indonesian 200 200 (100.00%) id=200 it Italian 200 200 (100.00%) it=200 ja Japanese 200 200 (100.00%) ja=200 kn Kannada 200 200 (100.00%) kn=200 ko Korean 200 200 (100.00%) ko=200 mk Macedonian 200 200 (100.00%) mk=200 ml Malayalam 200 200 (100.00%) ml=200
languages # precisions items mr Marathi 200 200 (100.00%) mr=200 ne Nepali 200 200 (100.00%) ne=200 nl Dutch 200 200 (100.00%) nl=200 no Norwegian 200 199 (99.50%) da=1, no=199 pa Punjabi 200 200 (100.00%) pa=200 pl Polish 200 200 (100.00%) pl=200 pt Portuguese 200 200 (100.00%) pt=200 ro Romanian 200 200 (100.00%) ro=200 ru Russian 200 200 (100.00%) ru=200 sk Slovak 200 200 (100.00%) sk=200 so Somali 200 200 (100.00%) so=200 sq Albanian 200 200 (100.00%) sq=200 sv Swedish 200 200 (100.00%) sv=200 sw Swahili 200 200 (100.00%) sw=200 ta Tamil 200 200 (100.00%) ta=200 te Telugu 200 200 (100.00%) te=200 th Thai 200 200 (100.00%) th=200 tl Tagalog 200 200 (100.00%) tl=200 tr Turkish 200 200 (100.00%) tr=200 uk Ukrainian 200 200 (100.00%) uk=200 ur Urdu 200 200 (100.00%) ur=200 vi Vietnamese 200 200 (100.00%) vi=200 zh-cn Simplified Chinese 200 200 (100.00%) zh-cn=200 zh-tw Traditional Chinese 200 200 (100.00%) zh-tw=200 sum 9800 9777 (99.77%)
Language Detection with Naive Bayes Classifies documents into “language” categories Categories: English, Japanese, Chinese, … Updates the posterior probabilities of categories by feature probabilities in each category 𝑝 𝐶k 𝑋 (m+1) ∝ 𝑝 𝐶k 𝑋 m ⋅ 𝑝 𝑋𝑖 𝐶𝑘 where
𝐶k:category, 𝑋:document, 𝑋𝑖:feature of document Terminates detection process if the maximum probability(normalized) is over 0.99999 Early termination for perfomance
Features of Language Detection Character n-gram
To be exact, “Unicode’s codepoint n-gram” Much less than the size of words Separator of words □ T h i s □ T h i s ←1-gram □T Th hi is s□ ←2-gram □Th Thi his is□ ←3-gram
How to detect the text’s language Each language has the peculiar characters and spelling rule.
The accented “é” is used in Spanish, Italian and so on, and not used in English in principle. The word that starts with “Z” is often used in German and rarely used in English. The word that starts with “C” and contains spell “Th” are used in English and not used in German. Accumulates the probabilities assigned to these features in
given text, so the guessed language is obtained as one that has the maximum probability. □C □L □Z Th English 0.75 0.47 0.02 0.74 German 0.10 0.37 0.53 0.03 French 0.38 0.69 0.01 0.01
Improvement for Naive Detection The above naive algorithm can detect only 90% precision. Not “practical” Very low precision for some languages Japanese, Traditional Chinese, Russian, Persian, ...
Bias and noise of training and test corpus Improvement
Noise filter Character normalization
(1) Bias of Characters Alphabet / Arabic / Devanagari
About 30 characters Kanji (Chinese character)
20000 characters over! 1000 times as much as Alphabets
Kanji has “zero frequency problem”
Can’t detect language of “谢谢”(Simplified Chinese) This character isn’t used on Wikipedia.
Name Kanji (uneven frequency)
Normalization with “Joyo Kanji” Classifies “similar frequency Kanji” and normalizes
each cluster into a representative Kanji. (1) Clustering by K-means (2) Classification by “Joyo Kanji” Joyo Kanji (
常用漢字: regularly used Kanji) Simplified Chinese: “
现代汉语常用字表”(3500 characters) Traditional Chinese: the first standard of Big5 (5401 characters). It includes “常用国字標準字体表” (4808 characters) Japanese: Joyo Kanji(2136 characters) + the first standard of JIS X 0208 (2965 characters) = 2998 characters 130 clusters Each language has about 50 classes.
(2) Noise of Corpus Removes the language-independent characters
Numeric figures, symbols, URLs and mail addresses Latin character noise in non-Latin text
Alphabets often occur in also non-Latin text. Java, WWW, US and so on
Remove all Latin-characters if their rate is less than 20%. Latin character noise in Latin text
Acronyms, person’s names and place names don’t represent feature of languages. UNESCO, “New York” in French text
Person’s name has a various language feature (e.g. Mc- = Gaelic).
Removes all-capital words Reduces the effect of local features by the feature-sampling
Normalization of Arabic Character All Persian texts were detected as Arabic!
Persian and Arabic belong to different language families, so it ought to be easy to discriminate them. A high-frequency character “yeh” is assigned to
different codes in training and test corpora respectively. In the training corpus (Wikipedia), it is assigned to “ی” (¥u06cc, Farsi yeh). In the test corpus (News), it is “ي” (¥u064a, Arabic yeh). Cause: Arabic character-code CP-1256 don’t has the character mapped to ¥u06cc, so it is substituted to ¥u064a in a general way. Normalizes ¥u06cc(Farsi yeh) into ¥u064a(Arabic yeh)
All Persian texts are detected correctly.
Conclusion We developed the language detection
library for Java. 49 languages can be detected in 99.8% precision. Our next product will use it (search by language). 90% is easy. But 99% over is practical.
Ideal: Answer from the novel beautiful theory Real: Unrefined steady way all along
Open Issues Short text (e.g. twitter)
Arabic vowel signs
Text in more than one language
Source code in text
References [Habash 2009] Introduction to Arabic Natual Language Processing