Extracting Key Semantic Terms from Chinese Speech Query for Web Searches
36901Authors
nan
Abstract
� This� paper� discusses� the� challenges� and� pro- poses� a� solution� to� performing� information� re- trieval�on�the�Web�using�Chinese�natural�language� speech� query.� The� main� contribution� of� this� re- search�is�in�devising�a�divide-and-conquer�strategy� to� alleviate� the� speech� recognition� errors.� It� uses� the�query�model�to�facilitate�the�extraction�of�main� core�semantic�string�(CSS)�from�the�Chinese�natu- ral�language�speech�query.�It�then�breaks�the�CSS� into� basic� components� corresponding� to� phrases,� and� uses� a� multi-tier� strategy� to� map� the� basic� components� to� known� phrases� in� order� to� further� eliminate�the�errors.�The�resulting�system�has�been� found�to�be�effective.� 1�
Keywords
Full Content
Extracting�Key�Semantic�Terms�from�Chinese�Speech�Query�for�Web� Searches� � Gang�WANG� � National�University�of� Singapore� wanggang_sh@hotmail.com� � Tat-Seng�CHUA��� � National�University�of�Singa- pore� chuats@comp.nus.edu.sg� � Yong-Cheng�WANG� � Shanghai�Jiao�Tong�Univer- sity,�China,�200030� ycwang@mail.sjtu.edu.cn� Abstract� This� paper� discusses� the� challenges� and� pro- poses� a� solution� to� performing� information� re- trieval�on�the�Web�using�Chinese�natural�language� speech� query.� The� main� contribution� of� this� re- search�is�in�devising�a�divide-and-conquer�strategy� to� alleviate� the� speech� recognition� errors.� It� uses� the�query�model�to�facilitate�the�extraction�of�main� core�semantic�string�(CSS)�from�the�Chinese�natu- ral�language�speech�query.�It�then�breaks�the�CSS� into� basic� components� corresponding� to� phrases,� and� uses� a� multi-tier� strategy� to� map� the� basic� components� to� known� phrases� in� order� to� further� eliminate�the�errors.�The�resulting�system�has�been� found�to�be�effective.� 1� Introduction�� We� are� entering� an� information� era,� where� infor- mation�has�become�one�of�the�major�resources�in� our�daily�activities.�With�its�wide�spread�adoption,� Internet�has�become�the�largest�information�wealth� for� all� to� share.� Currently,� most� (Chinese)� search� engines� can� only� support� term-based� information� retrieval,�where�the�users�are�required�to�enter�the� queries�directly�through�keyboards�in�front�of�the� computer.� However,� there� is� a� large� segment� of� population�in�China�and�the�rest�of�the�world�who� are�illiterate�and�do�not�have�the�skills�to�use�the� computer.�They�are�thus�unable�to�take�advantage� of�the�vast�amount�of�freely�available�information.� Since� almost� every� person� can� speak� and� under- stand�spoken�language,�the�research�on�“(Chinese)� natural� language� speech� query� retrieval”� would� enable�average�persons�to�access�information�using� the�current�search�engines�without�the�need�to�learn� special�computer�skills�or�training.�They�can�sim- ply� access� the� search� engine� using� common� de- vices� that� they� are� familiar� with� such� as� the� telephone,�PDA�and�so�on.� In�order�to� implement�a�speech-based�informa- tion� retrieval� system,� one� of� the� most� important� challenges�is�how�to�obtain�the�correct�query�terms� from�the�spoken�natural�language�query�that�con- vey�the�main�semantics�of�the�query.�This�requires� the�integration�of�natural�language�query�process- ing�and�speech�recognition�research.� Natural�language�query�processing�has�been�an� active�area� of�research� for� many� years� and� many� techniques� have� been� developed� (Jacobs� and� Rau1993;�Kupie,�1993;�Strzalkowski,�1999;�Yu�et� al,�1999).�Most�of�these�techniques,�however,�focus� only�on�written�language,�with�few�devoted�to�the� study�of�spoken�language�query�processing.� Speech� recognition� involves� the� conversion� of� acoustic�speech�signals�to�a�stream�of�text.�Because� of�the�complexity�of�human�vocal�tract,�the�speech� signals�being�observed�are�different,�even�for�mul- tiple�utterances�of�the�same�sequence�of�words�by� the�same�person�(Lee�et�al�1996).�Furthermore,�the� speech�signals�can�be�influenced�by�the�differences� across� different� speakers,� dialects,� transmission� distortions,� and� speaking� environments.� These� have� contributed� to� the� noise� and� variability� of� speech�signals.�As�one�of�the�main�sources�of�er- rors�in�Chinese�speech�recognition�come�from�sub- stitution� (Wang� 2002;� Zhou� 1997),� in� which� a� wrong�but�similar�sounding�term�is�used�in�place�of� the�correct�term,�confusion�matrix�has�been�used�to� record�confused�sound�pairs�in�an�attempt�to�elimi- nate� this� error.� Confusion� matrix� has� been� em- ployed� effectively� in� spoken� document� retrieval� (Singhal�et�al,�1999�and�Srinivasan�et�al�2000)�and� to�minimize�speech�recognition�errors�(Shen�et�al,� 1998).� However,� when� such� method� is� used� di- rectly�to�correct�speech�recognition�errors,�it�tends� to�bring�in� too� many� irrelevant�terms�(Ng� 2000).� Because� important� terms� in� a� long� document� are� often�repeated�several�times,�there�is�a�good�chance� that�such�terms�will�be�correctly�recognized�at�least� once�by�a�speech�recognition�engine�with�a�reason- able�level�of� word�recognition� rate.�Many�spoken� document�retrieval�(SDR)�systems�took�advantage� of�this�fact�in�reducing�the�speech�recognition�and� matching�errors�(Meng�et�al�2001;�Wang�et�al�2001;� Chen� et� al� 2001).� In� contrast� to� SDR,� very� little� work� has� been� done� on� Chinese� spoken� query� processing�(SQP),�which�is�the�use�of�spoken�que- ries�to�retrieval�textual�documents.�Moreover,�spo- ken�queries�in�SQP�tend�to�be�very�short�with�few� repeated�terms.� In� this� paper,� we� aim� to� integrate� the� spoken� language�and�natural�language�research�to�process� spoken�queries�with�speech�recognition�errors.�The� main�contribution�of�this�research�is�in�devising�a� divide-and-conquer�strategy�to�alleviate�the�speech� recognition� errors.� It� first� employs� the� Chinese� query� model� to� isolate� the� Core� Semantic� String� (CSS)� that� conveys� the� semantics� of� the� spoken� query.� It� then� breaks� the� CSS� into� basic� compo- nents�corresponding�to�phrases,�and�uses�a�multi- tier�strategy�to�map�the�basic�components�to�known� phrases�in�a�dictionary�in�order�to�further�eliminate� the�errors.� In�the�rest�of�this�paper,�an�overview�of�the�pro- posed�approach�is�introduced�in�Section�2.�Section� 3�describes�the�query�model,�while�Section�4�out- lines� the� use� of� multi-tier� approach� to� eliminate� errors�in�CSS.�Section�5�discusses�the�experimental� setup� and� results.� Finally,� Section� 6� contains� our� concluding�remarks.� 2� Overview�of�the�proposed�approach� There�are�many�challenges�in�supporting�surfing�of� Web�by�speech�queries.�One�of�the�main�challenges� is�that�the�current�speech�recognition�technology�is� not�very�good,�especially�for�average�users�that�do� not�have�any�speech�trainings.�For�such�unlimited� user� group,� the� speech� recognition� engine� could� achieve�an�accuracy�of�less�than�50%.�Because�of� this,�the�key� phrases� we� derived� from� the� speech� query�could�be�in�error�or�missing�the�main�seman- tic� of� the� query� altogether.�This� would� affect� the� effectiveness�of�the�resulting�system�tremendously.� Given�the�speech-to-text�output�with�errors,�the� key�issue�is�on�how�to�analyze�the�query�in�order�to� grasp�the�Core�Semantic�String�(CSS)�as�accurately� as� possible.� CSS� is� defined� as� the� key� term� se- quence�in�the�query�that�conveys�the�main�seman- tics� of� the� query.� For� example,� given� the� query:� “ �������� � �� ���� � �!" #$�%& ('�) ”�(Please�tell� me�the�information�on�how�the�U.S.�separates�the� most-favored-nation� status� from� human� rights� is- sue�in�china).�The�CSS�in�the�query�is�underlined.� We�can�segment�the�CSS�into�several�basic�com- ponents�that� correspond� to� key� concepts� such� as:� * � (U.S.),� � (China),� �+ � (human� rights�issue),� !" #�$ �(the�most-favored-nation� status)�and� %�& �(separate).� Because� of� the� difficulty� in� handling� speech� recognition� errors� involving�multiple� segments� of� CSSs,�we�limit�our�research�to�queries�that�contain� only�one�CSS�string.�However,�we�allow�a�CSS�to� include� multiple�basic�components� as�depicted� in� the�above�example.�This�is�reasonable�as�most�que- ries�posed�by�the�users�on�the�Web�tend�to�be�short� with�only�a�few�characters�(Pu�2000).� Thus� the� accurate� extraction� of� CSS� and� …
Content truncated to 500 words
Labeled Text
� This� paper� discusses� the� challenges� and� proposes� a� solution� to� performing� information� retrieval�on�the�Web�using�Chinese�natural�language� speech� query.� The� main� contribution� of� this� research�is�in�devising�a�divide-and-conquer�strategy� to� alleviate� the� speech� recognition� errors.� It� uses� the�query�model�to�facilitate�the�extraction�of�main� core�semantic�string�(CSS)�from�the�Chinese�natural�language�speech�query.�It�then�breaks�the�CSS� into� basic� components� corresponding� to� phrases,� and� uses� a� multi-tier� strategy� to� map� the� basic� components� to� known� phrases� in� order� to� further� eliminate�the�errors.�The�resulting�system�has�been� found�to�be�effective.� 1�
Publication Details
- Paper ID:
- 36901
- Year:
- 2003
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan