/UniDic2UD

Tokenizer POS-tagger Lemmatizer and Dependency-parser for modern and contemporary Japanese

Primary LanguagePythonMIT LicenseMIT

Current PyPI packages

UniDic2UD

Tokenizer, POS-tagger, lemmatizer, and dependency-parser for modern and contemporary Japanese, working on Universal Dependencies.

Basic usage

>>> import unidic2ud
>>> nlp=unidic2ud.load("kindai")
>>> s=nlp("其國を治めんと欲する者は先づ其家を齊ふ")
>>> print(s)
# text = 其國を治めんと欲する者は先づ其家を齊ふ
1		其の	DET	連体詞	_	2	det	_	SpaceAfter=No|Translit=ソノ
2			NOUN	名詞-普通名詞-一般	_	4	obj	_	SpaceAfter=No|Translit=クニ
3			ADP	助詞-格助詞	_	2	case	_	SpaceAfter=No|Translit=
4	治め	収める	VERB	動詞-一般	_	7	advcl	_	SpaceAfter=No|Translit=オサメ
5			AUX	助動詞	_	4	aux	_	SpaceAfter=No|Translit=
6			ADP	助詞-格助詞	_	4	case	_	SpaceAfter=No|Translit=
7	欲する	欲する	VERB	動詞-一般	_	8	acl	_	SpaceAfter=No|Translit=ホッスル
8			NOUN	名詞-普通名詞-一般	_	14	nsubj	_	SpaceAfter=No|Translit=モノ
9			ADP	助詞-係助詞	_	8	case	_	SpaceAfter=No|Translit=
10	先づ	先ず	ADV	副詞	_	14	advmod	_	SpaceAfter=No|Translit=マヅ
11		其の	DET	連体詞	_	12	det	_	SpaceAfter=No|Translit=ソノ
12			NOUN	名詞-普通名詞-一般	_	14	obj	_	SpaceAfter=No|Translit=ウチ
13			ADP	助詞-格助詞	_	12	case	_	SpaceAfter=No|Translit=
14	齊ふ	整える	VERB	動詞-一般	_	0	root	_	SpaceAfter=No|Translit=トトノフ

>>> t=s[7]
>>> print(t.id,t.form,t.lemma,t.upos,t.xpos,t.feats,t.head.id,t.deprel,t.deps,t.misc)
7 欲する 欲する VERB 動詞-一般 _ 8 acl _ SpaceAfter=No|Translit=ホッスル

>>> print(s.to_tree())
     <══╗         det(決定詞)
     ═╗═╝<obj(目的語)
     <╝   ║       case(格表示)
  治め ═╗═╗═╝<advcl(連用修飾節)
     <╝ ║   ║     aux(動詞補助成分)
     <══╝   ║     case(格表示)
欲する ═══════╝<acl(連体修飾節)
     ═╗═══════╝<nsubj(主語)
     <╝         ║ case(格表示)
  先づ <══════╗   ║ advmod(連用修飾語)
     <══╗   ║   ║ det(決定詞)
     ═╗═╝<╗ ║   ║ obj(目的語)
     <╝   ║ ║   ║ case(格表示)
  齊ふ ═════╝═╝═══╝ root()

>>> f=open("trial.svg","w")
>>> f.write(s.to_svg())
>>> f.close()

trial.svg

unidic2ud.load(UniDic,UDPipe) loads a natural language processor pipeline, which uses UniDic for tokenizer POS-tagger and lemmatizer, then uses UDPipe for dependency-parser. The default UDPipe is UDPipe="japanese-modern". Available UniDic options are:

unidic2ud.UniDic2UDEntry.to_tree() has an option to_tree(BoxDrawingWidth=2) for old terminals, whose Box Drawing characters are "fullwidth".

You can simply use unidic2ud on the command line:

echo 其國を治めんと欲する者は先づ其家を齊ふ | unidic2ud -U kindai

CaboCha emulator usage

>>> import unidic2ud.cabocha as CaboCha
>>> c=CaboCha.Parser("kindai")
>>> s=c.parse("其國を治めんと欲する者は先づ其家を齊ふ")
>>> print(s.toString(CaboCha.FORMAT_TREE_LATTICE))
  -D
  國を-D
治めんと-D
    欲する-D
        者は-------D
          先づ-----D
              -D |
              家を-D
                齊ふ
EOS
* 0 1D 0/0 0.000000
	連体詞,*,*,*,*,*,其の,ソノ,*,DET	O	1<-det-2
* 1 2D 0/1 0.000000
	名詞,普通名詞,一般,*,*,*,,クニ,*,NOUN	O	2<-obj-4
	助詞,格助詞,*,*,*,*,,,*,ADP	O	3<-case-2
* 2 3D 0/1 0.000000
治め	動詞,一般,*,*,*,*,収める,オサメ,*,VERB	O	4<-advcl-7
	助動詞,*,*,*,*,*,,,*,AUX	O	5<-aux-4
	助詞,格助詞,*,*,*,*,,,*,ADP	O	6<-case-4
* 3 4D 0/0 0.000000
欲する	動詞,一般,*,*,*,*,欲する,ホッスル,*,VERB	O	7<-acl-8
* 4 8D 0/1 0.000000
	名詞,普通名詞,一般,*,*,*,,モノ,*,NOUN	O	8<-nsubj-14
	助詞,係助詞,*,*,*,*,,,*,ADP	O	9<-case-8
* 5 8D 0/0 0.000000
先づ	副詞,*,*,*,*,*,先ず,マヅ,*,ADV	O	10<-advmod-14
* 6 7D 0/0 0.000000
	連体詞,*,*,*,*,*,其の,ソノ,*,DET	O	11<-det-12
* 7 8D 0/1 0.000000
	名詞,普通名詞,一般,*,*,*,,ウチ,*,NOUN	O	12<-obj-14
	助詞,格助詞,*,*,*,*,,,*,ADP	O	13<-case-12
* 8 -1D 0/0 0.000000
齊ふ	動詞,一般,*,*,*,*,整える,トトノフ,*,VERB	O	14<-root
EOS
>>> for c in [s.chunk(i) for i in range(s.chunk_size())]:
...   if c.link>=0:
...     print(c,"->",s.chunk(c.link))
...
 -> 國を
國を -> 治めんと
治めんと -> 欲する
欲する -> 者は
者は -> 齊ふ
先づ -> 齊ふ
 -> 家を
家を -> 齊ふ

CaboCha.Parser(UniDic) is an alias for unidic2ud.load(UniDic,UDPipe="japanese-modern"), and its default is UniDic=None. CaboCha.Tree.toString(format) has five available formats:

  • CaboCha.FORMAT_TREE: tree (numbered as 0)
  • CaboCha.FORMAT_LATTICE: lattice (numbered as 1)
  • CaboCha.FORMAT_TREE_LATTICE: tree + lattice (numbered as 2)
  • CaboCha.FORMAT_XML: XML (numbered as 3)
  • CaboCha.FORMAT_CONLL: Universal Dependencies CoNLL-U (numbered as 4)

You can simply use udcabocha on the command line:

echo 其國を治めんと欲する者は先づ其家を齊ふ | udcabocha -U kindai -f 2

-U UniDic specifies UniDic. -f format specifies the output format in 0 to 4 above (default is -f 0) and in 5 to 8 below:

dot.png

Try notebook for Google Colaboratory.

Usage via spaCy

If you have already installed spaCy 2.1.0 or later, you can use UniDic via spaCy Language pipeline.

>>> import unidic2ud.spacy
>>> nlp=unidic2ud.spacy.load("kindai")
>>> d=nlp("其國を治めんと欲する者は先づ其家を齊ふ")
>>> print(unidic2ud.spacy.to_conllu(d))
# text = 其國を治めんと欲する者は先づ其家を齊ふ
1		其の	DET	連体詞	_	2	det	_	SpaceAfter=No|Translit=ソノ
2			NOUN	名詞-普通名詞-一般	_	4	obj	_	SpaceAfter=No|Translit=クニ
3			ADP	助詞-格助詞	_	2	case	_	SpaceAfter=No|Translit=
4	治め	収める	VERB	動詞-一般	_	7	advcl	_	SpaceAfter=No|Translit=オサメ
5			AUX	助動詞	_	4	aux	_	SpaceAfter=No|Translit=
6			ADP	助詞-格助詞	_	4	case	_	SpaceAfter=No|Translit=
7	欲する	欲する	VERB	動詞-一般	_	8	acl	_	SpaceAfter=No|Translit=ホッスル
8			NOUN	名詞-普通名詞-一般	_	14	nsubj	_	SpaceAfter=No|Translit=モノ
9			ADP	助詞-係助詞	_	8	case	_	SpaceAfter=No|Translit=
10	先づ	先ず	ADV	副詞	_	14	advmod	_	SpaceAfter=No|Translit=マヅ
11		其の	DET	連体詞	_	12	det	_	SpaceAfter=No|Translit=ソノ
12			NOUN	名詞-普通名詞-一般	_	14	obj	_	SpaceAfter=No|Translit=ウチ
13			ADP	助詞-格助詞	_	12	case	_	SpaceAfter=No|Translit=
14	齊ふ	整える	VERB	動詞-一般	_	0	root	_	SpaceAfter=No|Translit=トトノフ

>>> t=d[6]
>>> print(t.i+1,t.orth_,t.lemma_,t.pos_,t.tag_,t.head.i+1,t.dep_,t.whitespace_,t.norm_)
7 欲する 欲する VERB 動詞-一般 8 acl  ホッスル

>>> from deplacy.deprelja import deprelja
>>> for b in unidic2ud.spacy.bunsetu_spans(d):
...   for t in b.lefts:
...     print(unidic2ud.spacy.bunsetu_span(t),"->",b,"("+deprelja[t.dep_]+")")
...
 -> 國を (決定詞)
國を -> 治めんと (目的語)
治めんと -> 欲する (連用修飾節)
欲する -> 者は (連体修飾節)
 -> 家を (決定詞)
者は -> 齊ふ (主語)
先づ -> 齊ふ (連用修飾語)
家を -> 齊ふ (目的語)

unidic2ud.spacy.load(UniDic,parser) loads a spaCy pipeline, which uses UniDic for tokenizer POS-tagger and lemmatizer (as shown above), then uses parser for dependency-parser. The default parser is parser="japanese-modern" and available options are:

Installation for Linux

Tar-ball is available for Linux, and is installed by default when you use pip:

pip install unidic2ud

By default installation, UniDic is invoked through Web APIs. If you want to invoke them locally and faster, you can download UniDic which you use just as follows:

python -m unidic2ud download kindai
python -m unidic2ud dictlist

Licenses of dictionaries and models are: GPL/LGPL/BSD for gendai and spoken; CC BY-NC-SA 4.0 for others.

Installation for Cygwin

Make sure to get gcc-g++ python37-pip python37-devel packages, and then:

pip3.7 install unidic2ud

Use python3.7 command in Cygwin instead of python.

Installation for Jupyter Notebook (Google Colaboratory)

!pip install unidic2ud

Benchmarks

Results of 舞姬/雪國/荒野より-Benchmarks

舞姬 LAS MLAS BLEX
UniDic="kindai" 81.13 70.37 77.78
UniDic="qkana" 79.25 70.37 77.78
UniDic="kinsei" 72.22 60.71 64.29
雪國 LAS MLAS BLEX
UniDic="qkana" 89.29 85.71 81.63
UniDic="kinsei" 89.29 85.71 77.55
UniDic="kindai" 84.96 81.63 77.55
荒野より LAS MLAS BLEX
UniDic="kindai" 76.44 61.54 53.85
UniDic="qkana" 75.39 61.54 53.85
UniDic="kinsei" 71.88 58.97 51.28

Author

Koichi Yasuoka (安岡孝一)

References