This article explains step by step how to implement full-text search on a set of documents written in Chinese, Korean and Japanese languages (CJK).
About CJK languages
CJK languages have more than 40,000 characters. Most of them are Chinese. Sometimes you can see acronym CJKV. “V” here stands for the Vietnamese language.
CJK characters include:
- For the Chinese language: hànzì – traditional Chinese characters; Bopomofo – Chinese Phonetic Alphabet; Pinyin – Romanization of Chinese language (a concept close to the concept of transliteration).
- For the Japanese language: Hiragana – Japanese syllabary; Katakana – Japanese syllabary; Arabic numerals.
- For the Korean Language: Hangul (Korean alphabet)
In addition, each language has a set of hieroglyphic keys (radicals), which act as a grouping elements to search for characters in the dictionary or as a semantic elements that define the meaning of the characters following the key.
To display text in CJK languages you can use the following encodings: Big5, EUC-JP, EUC-KR, ISO 2022-JP, KS C 5861, Shift-JIS, Unicode, etc. For CJK-language alphabets there are such Unicode blocks (http://www.unicode.org/Public/UNIDATA/Blocks.txt):
|1100 .. 11FF||Hangul Jamo||A single character out of a syllable in the Korean Hangul alphabet. Letters Jamo used to form the syllables Hangul|
|2E80 .. 2EFF||CJK Radicals Supplement||Key (radical) – an element of the hieroglyphic alphabet, which allows grouping of words or acts as a semantic element that defines the meaning of the following characters.|
|2F00 .. 2FDF||Kangxi Radicals||list of keys Kangxi adopted in Japan, Korea, Taiwan, traditionally includes 214 characters|
|3000 .. 303F||CJK Symbols and Punctuation||Ideographic characters and punctuation|
|3040 .. 309F||Hiragana||Japanese syllabary|
|30A0 .. 30FF||Katakana||Japanese syllabary|
|3100 .. 312F||Bopomofo||Chinese Phonetic Alphabet|
|3130 .. 318F||Hangul Compatibility Jamo|
|3190 .. 319F||Kanbun Camboon or kanbun||One of the written languages of medieval Japan|
|31A0 .. 31BF||Bopomofo Extended|
|31C0 .. 31EF||CJK Strokes simple features (elements) characters|
|31F0 .. 31FF||Katakana Phonetic Extensions|
|3200 .. 32FF||Enclosed CJK Letters and Months||CJK letters and months in circles|
|3300 .. 33FF||CJK Compatibility|
|3400 .. 4DBF||CJK Unified Ideographs Extension||A CJK Ideographs|
|4DC0 .. 4DFF||Yijing Hexagram Symbols|
|4E00 .. 9FFF||CJK Unified Ideographs||Ideographs – written sign, conditional image or picture, is not the appropriate speech sounds, and whole word|
|A000 .. A48F||Yi Syllables Yi language||The language of the province of South Sichuan|
|A490 .. A4CF||Yi Radicals|
|AC00 .. D7AF||Hangul Syllables Syllables Hangul|
|D7B0 .. D7FF||Hangul Jamo Extended-B|
|20000 .. 2A6DF||CJK Unified Ideographs Extension B|
|2A700 .. 2B73F||CJK Unified Ideographs Extension C|
|2F800 .. 2FA1F||CJK Compatibility Ideographs Supplement|
Note that the Arabic numerals, which can be used in CJK texts, correspond widespace character codes (see section FFF0 .. FFFF; Specials).
You can see here http://www.utf8-chartable.de/ how certain characters look.
How to tell Manticore Search that your document has CJK characters?
Manticore Search filters the texts at character level. Characters not accepted for tokenization are considered invalid and replace with whitespace, which acts as separator. By default, only english and russian characters are tokenized (along with underscore and letters).
CJK languages feature characters that can form unsegmented texts. For these types of characters, Manticore can index contiguous groups of these characters as n-grams.
In the index configuration we need to adjust 3 settings:
- charset_table – main parameter to describe the characters. Contains a table of symbols and rules for case folding.
- ngram_chars – description of characters needed to split CJK text to words using the N-gram model;
- Set the value ngram_len to 1. This enables the n-gram feature. Currently only 1-grams are supported ( a text “ABCDEF” [where A to F are in ngram_chars list] is indexed as “A B C D E F”).
How to create descriptions for the parameters charset_table and ngram_chars
Update: newer versions of Manticore Search introduce a charset_table alias containing all CJK characters needed as well as a Chinese ICU morphology processor. For working with CJK in latest version, read the following article.
Or in other words, how to explain Manticore Search which UTF-8 character codes belong to the family of CJK languages?
You can use the sets for blocks of language from charset_tables Sphinx’s wiki page or using the data in the table above and the rules set in charset_table to make your description of the options (see 1-3 above) for the characters and letters for CJK languages. Be careful and double check that all blocks of the character ranges that you need are included into Manticore Search index character description in configuration file. For example, if you would use character set range descriptions that you get on the link above for indexing documents containing Lisu or Vai languages, search will not work properly.
Pay special attention to setting the ngram_chars parameter correctly. When searching Manticore Search will not look into these characters as search matches.
An example of a index with CJK setup can be found at cjk_index_example.zip.