Skip to content
/ jsc Public

Joint source channel model for Japanese Kana Kanji conversion, Chinese pinyin input and CJE mixed input.

Notifications You must be signed in to change notification settings

yohokuno/jsc

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

45 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

JSC: Joint Souce Channel Model and Decoder

JSC is an implementation of joint source channel or joint n-gram model with monotonic decoder.

A Joint Source-Channel Model for Machine Transliteration, Li Haizhou, Zhang Min, Su Jian.
http://acl.ldc.upenn.edu/acl2004/main/pdf/121_pdf_2-col.pdf

It can be used for machine transliteration, Japanese kana-kanji conversion, Chinese pinyin input, English word segmentation or pronunciation inference.

Requirement

JSC requires Unix, gcc, python, and marisa-trie. If you want to use RPC server, you also need libevent.

marisa-trie 0.2.0 or later
http://code.google.com/p/marisa-trie/

libevent 2.0 or later
http://libevent.org/

Install

To install JSC, type these commands into your console.

$ ./waf configure [--prefix=INSTALL_DIRECTORY]
$ ./waf build
$ sudo ./waf install

Usage

jsc-decode

jsc-decode command convert source string into target string via joint source channel model. You can provide queries through standard input line by line.

options:
-d directory: specify data directory or prefix (default: ./)
-f format: specify format (segment [default], plain, debug)
-t table: specify table [romaji] mode (both [default], on, off)
-l: turn off sentence-beginning/ending label

jsc-build

jsc-build command build model files in binary format from n-gram file in text format.

options:
-d directory: specify data directory or prefix (default: ./)
-m model: specify model file name (default: ngram)
-t trie_num: specify trie number in marisa-trie (default: 3)
-r: build reverse model

jsc-server

jsc-server command provides RPC server via simple TCP protocol. You can provide queries through telnet command line by line.

options:
-d directory: specify data directory or prefix (default: ./)
-f format: specify format (segment [default], plain, debug)
-t table: specify table [romaji] mode (both [default], on, off)
-p port: specify port number (default: 40714)
-l: turn off sentence-beginning/ending label

Sample Applications

Japanese Kana Kanji Conversion

For Japanese Kana Kanji conversion, a model is provided at data/japanese directory. By default, both romaji and hiragana input are allowed.

$ ./build/jsc-decode -d data/japanese/
わたしのなまえはなかのです。
わたし の 名前 は 中野 です 。
arayurugenjitsuwosubetejibunnnohouhenejimagetanoda
あらゆる 現実 を 全て 自分 の ほう へ ネジ 曲げ た の だ

Japanese Pronunciation Inference

For Japanese pronunciation inference, a model is provided at data/japanese-reverse directory.

$ ./build/jsc-decode -d data/japanese-reverse/
魔理沙は大変なものを盗んでいきました
ま りさ は たいへん な もの を ぬす ん で い き ま し た

Chinese Pinyin Input

For Chinese Pinyin input, a model is provided at data/chinese/ directory.

$ ./build/jsc-decode -d data/chinese/
woaiziranyuyanchuli
我 爱 自然 语言 处理
zhejianshitagegehaibuzhidaone
这 件 事 她 哥哥 海部 知道 呢

Chinese Hanzi-to-Pinyin Conversion

For Chinese Hanzi-to-Pinyin Conversion, a model is provided at data/chinese-reverse/ directory.

$ ./build/jsc-decode -d data/chinese-reverse/
汉字拼音转换
hanzi pinyin zhuanhuan 

English word segmentation / automatic capitalization

For English input, a model is provided at data/english/ directory.

$ ./build/jsc-decode -d data/english/
alicewasbeginningtogetverytiredofsittingbyhersisteronthebank
Alice was beginning to get very tired of sitting by her sister on the bank
istandheretodayhumbledbythetaskbeforeusgratefulforthetrustyouhavebestowedmindfulofthesacrificesbornebyourancestors
I Stand here today humbled by the task before us grateful for the trust you have bestowed mindful of the sacrifices borne by our ancestors 

Mixed Input

For English/Japanese/Chinese mixed input, a model is provided at data/mixed/ directory. The language is detected automatically.

$ ./build/jsc-decode -d data/mixed/
thisisapencil
This is a pencil
kyouhayoitenkidesune
今日 は 良い 天気 です ね
woshizhongguoren
我 是 中国 人
thisistotemohaochi!
This is とても 好吃 !

Data Structure

Directories

Top directory contains these files and directories:

README.md   this file
build/      built by waf automatically
data/       model files
src/        source and header files for C++
tools/      command tools by C++
waf         waf build script
wscript     waf settings

File format

N-gram file should be SRILM format.

http://www.speech.sri.com/projects/srilm/

Target string and source string should be coupled with character '/'; e.g. "私/わたし"

Reference

Accuracy

LanguageF-scoreSize
Japanese0.93710MB
Chinese0.8959MB
Englishnot ready9MB
Mixednot ready27MB

Paper

Please refer this paper if you need.

Yoh Okuno and Shinsuke Mori, An Ensemble Model of Word-based and Character-based Models for Japanese and Chinese Input Method, Workshop on Advances in Text Input Methods, 2012.
http://yoh.okuno.name/pdf/wtim2012.pdf

About

Joint source channel model for Japanese Kana Kanji conversion, Chinese pinyin input and CJE mixed input.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages