commit | b56fa78a2fe44ac2851bae5bf4f4693a0644da7b | [log] [tgz] |
---|---|---|
author | cld2 <riesa@google.com> | Thu Aug 20 22:39:36 2015 |
committer | cld2 <riesa@google.com> | Thu Aug 20 22:39:36 2015 |
tree | 8084bc3703352e01415c8918c4567d9006b2374c | |
parent | 84b58a5d7690ebf05a91406f371ce00c3daf31c0 [diff] | |
parent | 49035cda2c57bde548e8e227168858e17ef05c3e [diff] |
Merge pull request #39 from kratorius/patch-1 Fix typo
Dick Sites (dsites@google.com)
2013.07.28
CLD2 probabilistically detects over 80 languages in Unicode UTF-8 text, either plain text or HTML/XML. Legacy encodings must be converted to valid UTF-8 by the caller. For mixed-language input, CLD2 returns the top three languages found and their approximate percentages of the total text bytes (e.g. 80% English and 20% French out of 1000 bytes of text means about 800 bytes of English and 200 bytes of French). Optionally, it also returns a vector of text spans with the language of each identified. This may be useful for applying different spelling-correction dictionaries or different machine translation requests to each span. The design target is web pages of at least 200 characters (about two sentences); CLD2 is not designed to do well on very short text, lists of proper names, part numbers, etc.
These 83 languages are detected:
Afrikaans Albanian Arabic Armenian Azerbaijani Basque Belarusian Bengali Bihari Bulgarian Catalan Cebuano Cherokee Croatian Czech Chinese Chinese_T Danish Dhivehi Dutch English Estonian Finnish French Galician Ganda Georgian German Greek Gujarati Haitian_Creole Hebrew Hindi Hmong Hungarian Icelandic Indonesian Inuktitut Irish Italian Javanese Japanese Kannada Khmer Kinyarwanda Korean Laothian Latvian Limbu Lithuanian Macedonian Malay Malayalam Maltese Marathi Nepali Norwegian Oriya Persian Polish Portuguese Punjabi Romanian Russian Scots_Gaelic Serbian Sinhalese Slovak Slovenian Spanish Swahili Swedish Syriac Tagalog Tamil Telugu Thai Turkish Ukrainian Urdu Vietnamese Welsh Yiddish.
Classification & Scoring. CLD2 is a Naïve Bayesian classifier, using one of three different token algorithms. For Unicode scripts such as Greek and Thai that map one-to-one to detected languages, the script defines the result. For the 80,000+ character Han script and its CJK combination with Hiragana, Katakana, and Hangul scripts, single letters (unigrams) are scored. For all other scripts, sequences of four letters (quadgrams) are scored.
Scoring is done exclusively on lowercased Unicode letters and marks, after expanding HTML entities &xyz; and after deleting digits, punctuation, and <tags>. Quadgram word beginnings and endings (indicated here by underscore) are explicitly used, so the word _look_ scores differently from the word-beginning _look or the mid-word look. Quadgram single-letter “words” are completely ignored. For each letter sequence, the scoring uses the 3-6 most likely languages and their quantized log probabilities. The training corpus is manually constructed from chosen web pages for each language, then augmented by careful automated scraping of over 100M additional web pages.
Several embellishments improve the basic algorithm:
Hints. Several hints can be supplied. Because these can be inaccurate on web pages, they are just hints -- they add a bias but do not force a specific language to be the detection result. The hints include:
Optimized for space and speed. The table-driven extraction of letter sequences and table-driven scoring is highly optimized for both space and speed, running about 10x faster than other detectors and covering over 70 languages in 1.8MB of x86 code and tables. The main quadgram lookup table consists of 256K four-byte entries, covering about 50 languages. Detection over the average web page of 30KB (half tags/digits/punctuation, half letters) takes roughly 1 msec on a current x86 processor.
CLD2 is an update of the original CLD, adding more languages, updating to Unicode 6.2 characters, improving scoring, and adding the optional output vector of labelled language spans.