ѧ
JOURNAL OF SOFTWARE
1999 10 12  Vol.10 No.12 1999



ĺѡ
塡ġ
ժҪǽʽﴦеҪ.ںУСλǴʣ˶ѡǿϵͳȽ.һֻھѭĺѡ񷽷,Ӵʻ﷨еϵ,﷨һ».ͨʹû氲ĿϽʵ,ѡĿŲԵȷΪ96.9%.
ؼʡ,ѭ,,,ѡ.
ͼšTP18
Word Sense Disambiguation of Spoken Chinese Using Neural Network
WANG Hai-fengGAO WenLI Sheng
(Department of Computer Science and Engineering Harbin Institute of Technology Harbin 150001)
AbstractSpoken Chinese analysis lies in the center of interactive speech processing system.The smallest meaningful unit in Chinese language is the word,so word sense disambiguation is the basis of spoken Chinese analysis.In this paper, the authors propose a novel method for spoken Chinese word sense disambiguation based on a simple recurrent network. This method provides a consistent processing strategy for syntax and semantics according to the internal logic between the word syntactic classification and semantic classification. Applied in the corpus for meeting schedule, this method achieves an accuracy of 96.9% in an open testing of word sense disambiguation.
Key wordsNeural network, simple recurrent network, spoken language analysis, spoken Chinese, word sense disambiguation.
ǿļеҪ֮һ.಻ͬص,磬еĲ󣨰ɡ֧ᡢظʲȣ﷨ԼԽûȷľӱ߽ȣ1,2.,̺ʹʶ.,иǿݴ³.
ͬһʻڲͬĻпܾвͬ﷨,ǴʵĶ.һʵ﷨һĻΨһȷ,ѡضﾳΪѡΨһȷ.ʵ﷨ǹԽṹĻ,ʻ㼶ѡ׼ȷ̶ֱӹϵĳɰ.
ΪҪ,оڽͬʱ,Ҳ˱Ҫĸ34.緽ǿѧϰõ³Զܵڶоߵӣ4.ѭ(simple recurrent network,ΪSRN)5ͨĵԪ(context units)ʹ߱˼ĵ.Wermter˽SRNڵȡõĽ4.Wermter˶﷨ѡԶ,﷨ϵ,ûпǵһ.
ȽSRN,ȻSRNĺѡ񷽷,һ﷨һ»,ʵó.
1ѭ
SRNElman,ʱԤ5.֮,ܶѧ߶SRNӦо4,6,7.ṹͼ1ʾ.

ͼ1
ѵ,Ĳ.,Ϳ԰Ϊһk+mԪ3ǰ,ֻǰkԪزǰһΪԼ.ͨ,SRNBP㷨ѵ.
ͨĲ,SRNʹýϴΧΪǰݣ5.
2ھѭĶѡ
2.1֪ʶ
 ڻﴦͻ,Ŀǰѧڴⷽоʱ,޶һȷӦ,ȽѶ,ȷӦüֵ.氲(meeting schedule)ĺӢ﷭8ݣ¼˹ڻ氲ŵĶԻ150,Ȼ˹ֵд¼ȫӦıϿԼ20 000(ԻиΪͷΪ).ִʴ,õ839ʵĺʱ.ͨ﷨ϵĴϸ,ȷ޹ص﷨26.Ե,򱳾й,,޶Ŀ,ȷҲص.Ĵӻ氲Իص,˰37弯.,ΪʱеĴ﷨,õôʵ.
һж,ÿԼ﷨.ͬһ﷨಻ͬʱ.1Ϊ.ľ̬Ͷֲ̬1.
1ķֲ

̬(Դʵͳƽ)̬(ͳƽ)
ٷֱ(%)ٷֱ(%)
ܴ83910 838
627.42 58023.8
﷨404.81 72115.9
546.42 23420.6

ɱ1Կ,ʵıȻ,ʹƵȴϸ,ѡбҪ.ֵָ,һͨôʵͨıϿΪʹõĴʵ޶. 
2.2﷨ѡ
ûжšŵԵķָ־,޷ԾΪλз.Ϊ,˻(dialog turn)9ݵĸ.һָһ˵˲ϵء˵һλ.ʵֵѡ㷨ԻΪλ,ѡʱֻʹûϢ.
ڽյpAp=ap1,ap2,...,apmʱ,ӦΪ:زHp=hp1,hp2,...,hpk,ĲBp=bp1,bp2,...,bpk,Op=op1,op2,...,opn.紦ѧϰ׶,ApӦTp=tp1,tp2,...,tpn.
Ĳֵ紦ǰһʱزֵ,ԣp2ʱ,Bp=Hp-1.p=1ʱ,b1i(-1,+1)ڵĳʼֵ.
Ĳk+mԪ,յpʱ,ΪCp=Bp,Ap=bp1,bp2,...,bpk,ap1,ap2,...,apm=cp1,cp2,...,cpk+m.
fԪص﷨༯W=w1,w2,...,wf(ϵͳf=26),m=f,n=m,Ԫֱ﷨༯WеԪһһӦ.زĲԪĿkݾȷ.
qʵĻS=s1,s2,...,sq,SȷǴΪR=r1,r2,...,rq,rpW(p=1,2,...,q).ڴsp(1pq),пܱǵļWp=Wd1,Wd2,...,Wdh,Wp2WWp٫,WpеԪɴʵиôʵдȷ.ڻ,spȷΪrp.
ϵͳspʱ,֮ӦAp=ap1,ap2,...,apmݵķapi·.ʵsp﷨wi(wiWp),api=1,api=0.ʱ,Tp=tp1,tp2,...,tpmݵķtpi,·.spȷΪwi(wi=rp),tpi=1,tpi=0.
ϵͳзΪѧϰ̺ѡ.ѧϰ,붯BP㷨10Ϊѧϰ㷨,ѵɾ˹עУԵĺԻ.ѡ,ʹѧϰ̵õ.pspAp,õӦXp=xp1,xp2,...,xpf.ֻ﷨ѡ,i=sp﷨Ϊwi.
2.3ѡ
ѡĹ﷨ѡ,иԪĸӦ.gļV=v1,v2,...,vg(ϵͳеg=37),m=g,n=m.
ѵõѡʱ,pspʱõYp=yp1,yp2,...,pyg.ֻѡ,spΪvj.
2.4 ﷨һ»
ֱSRN﷨ѡ,ֲһ.ͬһ﷨ֵͬʱ,Կ﷨(wi,vj)Ψһ־ÿһ.spu:Zsp=(wd1,ve1),(wd2,ve2),...,(wdu,veu).ж1iu,1dif,1eig.Ҷij,1i,ju,wdi=wdjveivej.1if,1jg,﷨(wi,vj)Zsp,(wi,vj)ڴspһµ,(wi,vj)ڴspǲһµ.ڴʵдʱ,Ϊspѡ﷨(wi,vj)spǲһµ,ζ﷨ѡһǴ,һ».
һȷѡ,ȱ֤﷨һµ.Ϊȷһ,ۺ﷨ѡXpѡYpΪspÿһ(wdi,vei)Zspۺ.wdiӦ﷨ѡΪxpdi,veiӦ﷨ѡΪypei,ǵļȨͣ
(1)
Ȩϵݾȷ.
(2)
ѡ(wdi,vei)Ϊsp.(wdi,vei)Zsp,ڴspȻһµ.
2.5ʵ
Ե2.1ĿԻ˹ע,õѵͲ.ȡ80Ϊѵ,70ΪŲԼ.ѵֱѵ﷨ѡ.Ȼֱѵõ﷨ѡѡв,ȷȨϵ,ԶѡۺϽв.ʱ,ֱѵͿŲԼõպͿŲԽ.Խʺ׼ȷ.

,Ҫѡܴ﷨Ͷѡзֱָ﷨ܴܴͶ.2г˵﷨ѡյĶѡ3ֽ.
2ѡʵ

ղԿŲ
(%)׼ȷ(%)(%)׼ȷ(%)
﷨ѡ89.498.384.097.5
ѡ91.598.285.797.1
ѡ93.098.387.396.9

ʵ,﷨ѡزԪĿΪ39,ѡزԪĿΪ50.ڱMMX166(32Mڴ),ѡٶԼΪ650/s.
ڽжѡʱ,Ȩϵȷ:
 ĳʼֵΪ0.2,õǰȨϵ=,ǰ׼ȷʦ=0,ı仯=0.01
 ݵǰĦֵжѡ,׼ȷʦ̣
 >,=,=̣
 =+ģ
 0.8תڣ
 =,.
ڷռͿż,ֱõֵΪ0.420.41,Ӧ׼ȷʼ2.
3
ͨ۲ͷʵ,ֿ¼ص.
(1) ڱ2,Ҫѡܴͬ,׼ȷʿɱԲǿɽϺõطӳ㷨.ѡʸ﷨ѡ,˵ѶӰС,﷨ϲĻ.ѡʺ﷨Զѡ,˵һ»ʵЧ,֤﷨һ,ʹ.
(2) ֵС0.5,﷨ѡһ,Ȩʱ,Ȩظ,Ҳһ˵˿Ľ﷨ɿ.
(3) ʹñĵķȡõЧԹΪԭ:³Ժݴ緽ȷʵܽϺõشͲ﷨Ŀ޶ȷӦЧ.
(4) ĵо޶Ӧչ.Ŀǰ,ƵġݵĿ޷ʵ֣1,2,4.ǿѧϰ֤˱ĵķչ,չʱֻռӹӦĿϣ¶ص弯дʵ伴.
ڱĵķ,¼Ҫ˵:
(1) ׼SRNֻʹϢǰ֮ǰϢ,δʹϢ,ΪʹϢ,һЩѧ˸Ľʩ7.ûвøĽʩǻԭ:Ϣʹʱտ,ף7еķʹԪ7ϢӰʵʱ.ʵ,ʹδʹϢ,׼ȷ൱,Ҫ.
(2) ѵвҪ,ѡķǶҲҪһ,ϵͳ.ĵķǶʹ˶ѡĸ,Ϣڲ״̬(Ĳ).,ǶʲμѵǱҪ.ϵͳĶѡٶȫԽܵ.
ĻSRNĺѡ񷽷Ӧ氲ŵĺӢ﷭ϵͳ,Ч.
*оõȻѧ𡢹863߿ƼĿ𡢹ҽ˲ŻйѧԺ˼ƻ.
߼飺,1971,ʿ,ҪоΪ,ѧ˹.
,1956,ʿ,ʿʦ,ҪоΪ˹ܣý弼.
,1943,,ʿʦ,ҪоΪ,ѧ˹.
ͨѶϵ:,100080,к֪·49ϣ
ߵλҵѧѧ빤ϵ150001
E-mail: i-haiwan@microsoft.com
ο
1Waibel Alex. Interactive translation of conversational speech. IEEE Computer, 1996,29(7):4148
2Kitano Hiroaki. Speech-to-Speech Translation: A Massively Parallel Memory-based Approach. Hingham, MA, Kluwer Academic Publishers, 1994
3Lavie A. GLR*: a robust grammar-focused parser for spontaneously spoken language Ph.D. Thesis. Pittsburgh, PA: Carnegie Mellon University, 1996
4Wermter S, Weber V. SCREEN: learning a flat syntactic and semantic spontaneous language analysis using artificial neural networks. Journal of Artificial Intelligence Research, 1997,6(1):3585
5Elman Jeffery L. Finding structure in time. Cognitive Science, 1990,14(2):179211
6Servan-Schrieber D, Cleeremans A, McClelland L. Graded state machines: the representation of temporal contingencies in simple recurrent networks. Machine Learning, 1991,7(23):161194
7ΰȨ,.SRNNĺıע.о뷢չ,1997,34(6):421426
(Liu Wei-quan, Zhong Yi-xin. Part-of-speech tagging with simple recurrent neural network. Computer Research and Development, 1997,34(6):421426)
8,,.ĺӢ﷭.:Ϊ,Ԭ.Թ.:廪ѧ,1997.219224
(Wang Hai-feng, Gao Wen, Li Sheng. Chinese-English spoken language translation in limited domain. In: Chen Li-wei, Yuan Qi eds. Language Engineering. Beijing: Tsinghua University Press, 1997. 219224)
9Ȼ.ѧ.ɳ:Ͻ,1988
(He Zi-ran. A Survey of Pragmatics. Changsha: Hu'nan Education Press, 1988)
10Rumelhart D E, McClelland J L. Parallel Distributed Processing, : Foundations. Cambridge, MA: MIT Press, 1986
1998-09-11յԭ,1998-12-01յ޸ĸ
