23

I'm using NLTK word_tokenizer to split a sentence into words.

I want to tokenize this sentence:

في_بيتنا كل شي لما تحتاجه يضيع ...ادور على شاحن فجأة يختفي ..لدرجة اني اسوي نفسي ادور شيء

The code I'm writing is: 

import re
import nltk

lex = u" في_بيتنا كل شي لما تحتاجه يضيع ...ادور على شاحن فجأة يختفي ..لدرجة اني اسوي نفسي ادور شيء"

wordsArray = nltk.word_tokenize(lex)
print " ".join(wordsArray)

The problem is that the word_tokenize function doesn't split by words. Instead, it splits by letters so that the output is:

"ف ي _ ب ي ت ن ا ك ل ش ي ل م ا ت ح ت ا ج ه ي ض ي ع ... ا د و ر ع ل ى ش ا ح ن ف ج أ ة ي خ ت ف ي .. ل د ر ج ة ا ن ي ا س و ي ن ف س ي ا د و ر ش ي ء"

Any ideas ?

What I've reached so far:

By trying the text in here, it appeared to be tokenized by letters. Also, however, other tokenizers tokenized it correctly. Does that mean that word_tokenize is for English only? Does that go for most of NLTK functions?

Garrett Hyde
  • 5,108
  • 7
  • 51
  • 53
Hady Elsahar
  • 2,061
  • 4
  • 28
  • 47
  • 1
    Does http://stackoverflow.com/questions/7386856/python-arabic-nlp help? (And a stemmer http://nltk.org/api/nltk.stem.html#module-nltk.stem.isri) – Jon Clements Oct 23 '12 at 17:12

2 Answers2

12

I always recommend using nltk.tokenize.wordpunct_tokenize. You can try out many of the NLTK tokenizers at http://text-processing.com/demo/tokenize/ and see for yourself.

Jacob
  • 4,174
  • 1
  • 23
  • 25
  • what is the difference between most of those tokenizers ? and does that mean that most NLTK functions won't work with arabic ? – Hady Elsahar Oct 24 '12 at 23:23
  • 5
    The TreebankWordTokenizer is trained on wall street journal text, which is ascii, so it never works well on unicode text. The PunctWordTokenizer is trained on more variety of text, but I find that it's less predictable than the rest of them, which use regular expressions, making them usable on any language, with predictable results. – Jacob Oct 25 '12 at 02:02
  • 1
    NLTK in general works just fine with arabic, and any unicode text, it's just that some models expect ascii, and therefore don't do well with unicode. – Jacob Oct 25 '12 at 02:03
2

this is the output i get with my code, but i recall unicode doesn't go well in python 2 and I used 3.5

nltk.word_tokenize('في_بيتنا كل شي لما تحتاجه يضيع ...ادور على شاحن فجأة يختفي ..لدرجة اني اسوي نفسي ادور شيء ')

['في_بيتنا', 'كل', 'شي', 'لما', 'تحتاجه', 'يضيع', '...', 'ادور', 'على', 'شاحن', 'فجأة', 'يختفي', '..لدرجة', 'اني', 'اسوي', 'نفسي', 'ادور', 'شيء']

Pradi KL
  • 658
  • 7
  • 13