Japanese NLP Library

1   Requirements

• Third Party Dependencies
  • Cabocha Japanese Morphological parser http://sourceforge.net/projects/cabocha/
• Python Dependencies
  • Python 2.6.* or above

1.1   Links

• All code at jProcessing Repo GitHub

• Documentation and HomePage and Sphinx

• PyPi Python Package

clone [email protected]:kevincobain2000/jProcessing.git

1.2   Install

In Terminal

bash$ python setup.py install

1.3   History

• 0.2

  • Sentiment Analysis of Japanese Text

• 0.1

  • Morphologically Tokenize Japanese Sentence
  • Kanji / Hiragana / Katakana to Romaji Converter
  • Edict Dictionary Search - borrowed
  • Edict Examples Search - incomplete
  • Sentence Similarity between two JP Sentences
  • Run Cabocha(ISO--8859-1 configured) in Python.
  • Longest Common String between Sentences
  • Kanji to Katakana Pronunciation
  • Hiragana, Katakana Chart Parser

2   Libraries and Modules

2.1   Tokenize jTokenize.py

In Python

>>> from jNlp.jTokenize import jTokenize
>>> input_sentence = u'私は彼を５日前、つまりこの前の金曜日に駅で見かけた'
>>> list_of_tokens = jTokenize(input_sentence)
>>> print list_of_tokens
>>> print '--'.join(list_of_tokens).encode('utf-8')

Returns:

... [u'\u79c1', u'\u306f', u'\u5f7c', u'\u3092', u'\uff15'...]
... 私--は--彼--を--５--日--前--、--つまり--この--前--の--金曜日--に--駅--で--見かけ--た

Katakana Pronunciation:

>>> print '--'.join(jReads(input_sentence)).encode('utf-8')
... ワタシ--ハ--カレ--ヲ--ゴ--ニチ--マエ--、--ツマリ--コノ--マエ--ノ--キンヨウビ--ニ--エキ--デ--ミカケ--タ

2.2   Cabocha jCabocha.py

Run Cabocha with original EUCJP or IS0-8859-1 configured encoding, with utf8 python

• If cabocha is configured as utf8 then see this http://nltk.googlecode.com/svn/trunk/doc/book-jp/ch12.html#cabocha

>>> from jNlp.jCabocha import cabocha
>>> print cabocha(input_sentence).encode('utf-8')

Output:

<sentence>
 <chunk id="0" link="8" rel="D" score="0.971639" head="0" func="1">
  <tok id="0" read="ワタシ" base="私" pos="名詞-代名詞-一般" ctype="" cform="" ne="O">私</tok>
  <tok id="1" read="ハ" base="は" pos="助詞-係助詞" ctype="" cform="" ne="O">は</tok>
 </chunk>
 <chunk id="1" link="2" rel="D" score="0.488672" head="2" func="3">
  <tok id="2" read="カレ" base="彼" pos="名詞-代名詞-一般" ctype="" cform="" ne="O">彼</tok>
  <tok id="3" read="ヲ" base="を" pos="助詞-格助詞-一般" ctype="" cform="" ne="O">を</tok>
 </chunk>
 <chunk id="2" link="8" rel="D" score="2.25834" head="6" func="6">
  <tok id="4" read="ゴ" base="５" pos="名詞-数" ctype="" cform="" ne="B-DATE">５</tok>
  <tok id="5" read="ニチ" base="日" pos="名詞-接尾-助数詞" ctype="" cform="" ne="I-DATE">日</tok>
  <tok id="6" read="マエ" base="前" pos="名詞-副詞可能" ctype="" cform="" ne="I-DATE">前</tok>
  <tok id="7" read="、" base="、" pos="記号-読点" ctype="" cform="" ne="O">、</tok>
 </chunk>

2.3   Kanji / Katakana /Hiragana to Tokenized Romaji jConvert.py

Uses data/katakanaChart.txt and parses the chart. See katakanaChart.

>>> from jNlp.jConvert import *
>>> input_sentence = u'気象庁が２１日午前４時４８分、発表した天気概況によると、'
>>> print ' '.join(tokenizedRomaji(input_sentence))
>>> print tokenizedRomaji(input_sentence)

...kisyoutyou ga ni ichi nichi gozen yon ji yon hachi hun  hapyou si ta tenki gaikyou ni yoru to
...[u'kisyoutyou', u'ga', u'ni', u'ichi', u'nichi', u'gozen',...]

katakanaChart.txt

• katakanaChartFile and hiraganaChartFile

2.4   Longest Common String Japanese jProcessing.py

On English Strings

>>> from jNlp.jProcessing import long_substr
>>> a = 'Once upon a time in Italy'
>>> b = 'Thre was a time in America'
>>> print long_substr(a, b)

Output

...a time in

On Japanese Strings

>>> a = u'これでアナタも冷え知らず'
>>> b = u'これでア冷え知らずナタも'
>>> print long_substr(a, b).encode('utf-8')

Output

...冷え知らず

2.5   Similarity between two sentences jProcessing.py

Uses MinHash by checking the overlap http://en.wikipedia.org/wiki/MinHash

>>> from jNlp.jProcessing import Similarities
>>> s = Similarities()
>>> a = 'There was'
>>> b = 'There is'
>>> print s.minhash(a,b)
...0.444444444444

>>> from jNlp.jProcessing import *
>>> a = u'これは何ですか？'
>>> b = u'これはわからないです'
>>> print s.minhash(' '.join(jTokenize(a)), ' '.join(jTokenize(b)))
...0.210526315789

3   Edict Japanese Dictionary Search with Example sentences

3.1   Sample Ouput Demo

3.2   Edict dictionary and example sentences parser.

This package uses the EDICT and KANJIDIC dictionary files. These files are the property of the Electronic Dictionary Research and Development Group , and are used in conformance with the Group's licence .

Edict Parser By Paul Goins, see edict_search.py Edict Example sentences Parse by query, Pulkit Kathuria, see edict_examples.py Edict examples pickle files are provided but latest example files can be downloaded from the links provided.

3.3   Charset

Two files

• utf8 Charset example file if not using src/jNlp/data/edict_examples

  To convert EUCJP/ISO-8859-1 to utf8

  iconv -f EUCJP -t UTF-8 path/to/edict_examples > path/to/save_with_utf-8
  

• ISO-8859-1 edict_dictionary file

Outputs example sentences for a query in Japanese only for ambiguous words.

3.4   Links

Latest Dictionary files can be downloaded here

3.5   edict_search.py

For all entries of sense definitions

>>> from jNlp.edict_search import *
>>> query = u'認める'
>>> edict_path = 'src/jNlp/data/edict-yy-mm-dd'
>>> kp = Parser(edict_path)
>>> for i, entry in enumerate(kp.search(query)):
...     print entry.to_string().encode('utf-8')

3.6   edict_examples.py

>>> from jNlp.edict_examples import *
>>> query = u'認める'
>>> edict_path = 'src/jNlp/data/edict-yy-mm-dd'
>>> edict_examples_path = 'src/jNlp/data/edict_examples'
>>> search_with_example(edict_path, edict_examples_path, query)

Output

認める

Sense (1) to recognize;
  EX:01 我々は彼の才能を*認*めている。We appreciate his talent.

Sense (2) to observe;
  EX:01 ｘ線写真で異状が*認*められます。We have detected an abnormality on your x-ray.

Sense (3) to admit;
  EX:01 母は私の計画をよいと*認*めた。Mother approved my plan.
  EX:02 母は決して私の結婚を*認*めないだろう。Mother will never approve of my marriage.
  EX:03 父は決して私の結婚を*認*めないだろう。Father will never approve of my marriage.
  EX:04 彼は女性の喫煙をいいものだと*認*めない。He doesn't approve of women smoking.
  ...

4   Sentiment Analysis Japanese Text

This section covers (1) Sentiment Analysis on Japanese text using Word Sense Disambiguation, Wordnet-jp (Japanese Word Net file name wnjpn-all.tab), SentiWordnet (English SentiWordNet file name SentiWordNet_3.*.txt).

4.1   Wordnet files download links

1. http://nlpwww.nict.go.jp/wn-ja/eng/downloads.html
2. http://sentiwordnet.isti.cnr.it/

4.2   How to Use

The following classifier is baseline, which works as simple mapping of Eng to Japanese using Wordnet and classify on polarity score using SentiWordnet.

• (Adnouns, nouns, verbs, .. all included)
• No WSD module on Japanese Sentence
• Uses word as its common sense for polarity score

>>> from jNlp.jSentiments import *
>>> jp_wn = '../../../../data/wnjpn-all.tab'
>>> en_swn = '../../../../data/SentiWordNet_3.0.0_20100908.txt'
>>> classifier = Sentiment()
>>> classifier.train(en_swn, jp_wn)
>>> text = u'監督、俳優、ストーリー、演出、全部最高！'
>>> print classifier.baseline(text)
...Pos Score = 0.625 Neg Score = 0.125
...Text is Positive

4.3   Japanese Word Polarity Score

>>> from jNlp.jSentiments import *
>>> jp_wn = '_dicts/wnjpn-all.tab' #path to Japanese Word Net
>>> en_swn = '_dicts/SentiWordNet_3.0.0_20100908.txt' #Path to SentiWordNet
>>> classifier = Sentiment()
>>> sentiwordnet, jpwordnet  = classifier.train(en_swn, jp_wn)
>>> positive_score = sentiwordnet[jpwordnet[u'全部']][0]
>>> negative_score = sentiwordnet[jpwordnet[u'全部']][1]
>>> print 'pos score = {0}, neg score = {1}'.format(positive_score, negative_score)
...pos score = 0.625, neg score = 0.0

5   Contacts

Author:	pulkit[at]jaist.ac.jp [change at with @]