Building a Large-Scale Annotated Chinese Corpus
6374Authors
nan
Abstract
�� In� this� paper� we� address� issues� related� to� building� a� large-scale� Chinese� corpus.� We� try� to� answer� four� questions:� (i)� how� to� speed� up� annotation,� (ii)� how� to� maintain� high� annotation� quality,� (iii)� for� what� purposes� is� the� corpus� applicable,� and� finally�(iv)�what�future�work�we�anticipate.��
Keywords
Full Content
� Building�a�Large-Scale�Annotated�Chinese�Corpus� � Nianwen�Xue� IRCS,�University�of�Pennsylvania�� Suite�400A,�3401�Walnut�Street�� Philadelphia,�PA�19104,�USA� xueniwen@linc.cis.upenn.edu� Fu-Dong�Chiou�and�Martha�Palmer� CIS,�University�of�Pennsylvania� 200�S�33rd�Street�� Philadelphia,�PA�19104,�USA� {chioufd,mpalmer}@linc.cis.upenn.edu� � Abstract�� In� this� paper� we� address� issues� related� to� building� a� large-scale� Chinese� corpus.� We� try� to� answer� four� questions:� (i)� how� to� speed� up� annotation,� (ii)� how� to� maintain� high� annotation� quality,� (iii)� for� what� purposes� is� the� corpus� applicable,� and� finally�(iv)�what�future�work�we�anticipate.�� Introduction� The� Penn� Chinese� Treebank� (CTB)� is� an� ongoing� project,� with� its� objective� being� to� create� a� segmented� Chinese� corpus� annotated� with�POS�tags�and�syntactic�brackets.�The�first� installment� of� the� project� (CTB-I)� consists� of� Xinhua� newswire� between� the� years� 1994� and� 1998,�totaling�100,000�words,�fully�segmented,� POS-tagged� and� syntactically� bracketed� and� it� has� been� released� to� the� public� via� the� Penn� Linguistic� Data� Consortium� (LDC).� The� preliminary� results� of� this� phase� of� the� project� have�been�reported�in�Xia�et�al�(2000).�Currently� the� second� installment� of� the� project,� the� 400,000-word�CTB-II�is�being�developed�and�is� expected�to�be�completed�early�in�the�year�2003.� CTB-II� will� follow� the� standards� set� up� in� the� segmentation� (Xia� 2000b),� POS� tagging� (Xia� 2000a)�and�bracketing�guidelines�(Xue�and�Xia� 2000)� and� it� will� use� articles� from� Peoples'� Daily,� Hong� Kong� newswire� and� material� translated�into�Chinese�from�other�languages�in� addition�to�the�Xinhua�newswire�used�in�CTB-I� �in�an�effort�to�diversify�the�sources.� The�availability�of�CTB-I�changed�our�approach� to�CTB-II�considerably.�Due�to�the�existence�of� CTB-I,� we� were� able� to� train� new� automatic� Chinese�language�processing�(CLP)�tools,�which� crucially� use� annotated� corpora� as� training� material.� These� tools� are� then� used� for� preprocessing�in�the�development�of�the�CTB-II.� We�also�developed�tools�to�control�the�quality�of� the�corpus.�In�this�paper,�we�will�address�three� issues� in� the� development� of� the� Chinese� Treebank:�annotation�speed,�annotation�accuracy� and� usability� of� the� corpus.� Specifically,� we� attempt�to�answer�four�questions:�(i)�how�do�we� speed�up�the�annotation�process,�(ii)�how�do�we� maintain� high� quality,� i.e.� annotation� accuracy� and� inter-annotator� consistency� during� the� annotation�process,�and�(iii)�for�what�purposes�is� the� corpus� applicable,� and� (iv)� what� are� our� future� plans?� Although� we� will� touch� upon� linguistic�problems�that�are�specific�to�Chinese,� we� believe� these� issues� are� general� enough� for� the�development�of�any�single�language�corpus.� 1� Annotation�Speed� There� are� three� main� factors� that� affect� the� annotation� speed�:� annotators’� background,� guideline� design� and� more� importantly,� the� availability� of� preprocessing� tools.� We� will� discuss� how� each� of� these� three� factors� affects� annotation�speed.�� 1.1� Annotator�Background� Even� with� the� best� sets� of� guidelines,� it� is� important� that� annotators� have� received� considerable� training� in� linguistics,� particularly� in�syntax.�In�both�the�segmentation/POS�tagging� phase� and� the� syntactic� bracketing� phase,� understanding� the� structure� of� the� sentences� is� essential�for�correct�annotation�with�reasonable� speed.�For�example,� � /de�is�assigned�two�part- of-speech�tags,�depending�on�where�it�occurs�in� the�sentence.�It�is�tagged�as�DEC�when�it�marks� the� end� of� the� preceding� modifying� clause� and� DEG� when� it� follows� a� nominal� phrase.� This� � distinction� is� useful� in� that� it� marks� different� relations�:� between� the� nominal� phrase� and� the� noun�head,�and�between�the�clause�and�the�noun� �head�respectively.�� 1.a.� ��� /NN� � /DEG� ��� /NN� ��������leader���������DE����������responsibility� ��������‘leader’s�responsibility’� ��b.� �� /NT� � /VV� � /DEC� � /NN� � ������recently�hold�������DE��������demonstration��/NN� �������‘recently�held�demonstration’� During� the� bracketing� phase,� the� modifying� clause�is�further�divided�into�relative�clauses�and� complement�(appositive)�clauses.�The�structures� of� these� two� types� of� clauses� are� different,� as� �illustrated�in�2:� 2.a.�(NP�(CP�(WHNP-1�*OP*)� ����������������������(CP��(IP��(NP-SBJ�(-NONE-�*T*-1))� ������������������������(VP�(NP-TMP� /NT)�����recently� �������������������������(VP� /hold)))���������������hold� ������������������������������� /DEC))�� ���������(NP� /NN� /NN))�����������������demonstration���� �‘recently�held�demonstration’� ��b.�(NP�(CP-APP�(IP�(NP-SBJ�(-NONE-�*pro*))� ��������������������������(VP�(PP� /P�����������������������������������to� �����(NP� /NN))����������nation� �����������������������������(VP� ! /VV)))���������responsible� ������������������������������� /DEC)� ���������(NP� "# /NN))��������������������������������������������attitude� �‘the�attitude�that�one�is�responsible�to�the�nation’� The� annotator� needs� to� make� his/her� own� judgment� as� to� whether� the� preceding� constituent� is� a� phrase� or� a� clause.� If� it� is� a� clause,� he� then� needs� to� decide� whether� it� is� a� complement�clause�or�a�relative�clause.�That�is� just�one�of�the�numerous�places�where�he�would� have�to�draw�upon�his�training�in�syntax�in�order� to� annotate� the� sentence� correctly� and� efficiently.�Although�it�is�hard�to�quantify�how� the� annotator's� background� can� affect� the� annotation�speed,�it�is�safe�to�assume�that�basic� training� in� syntax� is� very� important� for� his� performance.� � 1.2� How� Guideline� Design� can� Affect�Annotation�Speed� In� addition� to� the� annotator’s� background,� …
Content truncated to 500 words
Labeled Text
�� In� this� paper� we� address� issues� related� to� building� a� large-scale� Chinese� corpus.� We� try� to� answer� four� questions:� (i)� how� to� speed� up� annotation,� (ii)� how� to� maintain� high� annotation� quality,� (iii)� for� what� purposes� is� the� corpus� applicable,� and� finally�(iv)�what�future�work�we�anticipate.��
Publication Details
- Paper ID:
- 6374
- Year:
- 2002
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan