Question

What is the best tokenizer exist for processing Korean language?

I have tried CJKTokenizer in Solr4.0. It is doing the tokenization, but accuracy is very low.

Answer 1

POSTECH/K is a Korean Morphological Analyzer that is able to tokenize and POS tag Korean data without much effort. The software reports 90.7% on the corpus it was train and tested on (see http://nlp.postech.ac.kr/download/postag_k/9908_cljournal_gblee.pdf).

The POS tagging achieved 81% on the korean data of a multilingual corpus project i’ve been working on.

However, there’s a catch, you have to use windows to run the software. But I’ve a script to bypass that limitation, here’s the script:

#!/bin/bash -x
###############################################################################
## Sejong-Shell is a script to call POSTAG/SEJONG tagger on Unix Machine
## because POSTAG/Sejong is only usable in Korean Microsoft Windows environment
## the original POSTAG/Sejong can be downloaded from
## http://isoft.postech.ac.kr/Course/CS730b/2005/index.html
##
## Sejong-Shell is dependent on WINdows Emulator.
## The WINE program can be downloaded from
## http://www.winehq.org/download/
##
## The shell scripts accepts the input files from one directory and
## outputs the tagged files into another while retaining the filename
###############################################################################

cd <source-file_dir>
#<source_-ile_dir> is the directory that saves the textfiles that needs tagging
for file in `dir -d *`
do
    echo $file
    sudo cp <source-file_dir>/"$file" <POSTAG-Sejong_dir>/input.txt
    # <POSTAG-Sejong_dir> refers to the directory where the pos-tagger is saved
    wine start /Unix "$HOME/postagsejong/sjTaggerInteg.exe"
    sleep 30
    # This is necessary so that the file from the current loop won't be
    # overlapping with the next, do increase the time for sleep if the file
    # is large and needs more than 30 sec for POSTAG/Sejong to tag.
    sudo cp <POSTAG-Sejong_dir>/output.txt <target-file_dir>/"$file"
    # <target-file_dir> is where you want the output files to be stored
done

# Instead of the sleep command to prevent the overlap:
#   $sleep 30
# Alternatively, you can manually continue a loop with the following 
# command that continues a loop after a keystroke input:
#   $read -p "Press any key to continue…"

Note that the encoding for POSTECH/K is euc-kr, so if it’s utf8. you can use the following script to convert from utf8 to euc-kr.

#!/usr/bin/python # -*- coding: utf-8 -*-

'''
pre-sejong clean
'''

import codecs
import nltk
import os, sys, re, glob
from nltk.tokenize import RegexpTokenizer

reload(sys)
sys.setdefaultencoding('utf-8')

cwd = './gizaclean_ko' #os.getcwd()
wrd = './presejong_ko'

kr_sent_tokenizer = nltk.RegexpTokenizer(u'[^！？.?!]*[！？."www.*"]')


for infile in glob.glob(os.path.join(cwd, '*.txt')):
#   if infile == './extract_ko/singapore-sling.txt': continue
#   if infile == './extract_ko/ion-orchard.txt': continue
        print infile
        (PATH, FILENAME) = os.path.split(infile)
        reader = open(infile)
        writer = open(os.path.join(wrd, FILENAME).encode('euc-kr'),'w')
        for line in reader:
                para = []urlread = lambda url: urllib.urlopen(url).read()
                para.append (kr_sent_tokenizer.tokenize(unicode(line,'utf-8').strip()))
                for sent in para[0]:
            newsent = sent.replace(u'\xa0', ' '.encode('utf-8'))
            newsent2 = newsent.replace(u'\xe7', 'c'.encode('utf-8'))
            newsent3 = newsent2.replace(u'\xe9', 'e'.encode('utf-8'))
            newsent4 = newsent3.replace(u'\u2013', '-')
            newsent5 = newsent4.replace(u'\xa9', '(c)')
            newsent6 = newsent5.encode('euc-kr').strip()
            print newsent6
            writer.write(newsent6+'\n')     

(source for sejong-shell : Liling Tan. 2011. Building the foundation text for Nanyang Technological University – Multilingual Corpus (NTU-MC). Final year project. Singapore: Nanyang Technological University. pp. 44.)