ѧ
JOURNAL OF SOFTWARE
1999 2 No.2 1999



ṹȹϵԶȡ*
ܡǿ Ʋ
ժҪһڶṹֲص֪ʶʾ:ṹȹϵSPR(structure preference relation),ԲͬıSPRȡ:ϵԶֺԭʼıԶȡ,ͨͬʵ֤ЩĿԺʵ.,SPRӦǰ,һо.
ؼʡṹȹϵ,ṹ,֪ʶȡ,Ͽ.
ͼšTP18
Automatic Acquisition for Chinese Structure Preference Relations
ZHOU Qiang HUANG Chang-ning
Abstract A new ambiguity representation scheme SPR(structure preference relation) is proposed in this paper, which consists of useful quantitative distribution information for ambiguous structures. Two automatic acquisition algorithms, (1) acquired from treebank, (2) acquired from raw texts, are introduced, and some experimental results which prove the availability of the algorithms are also given. At last, some SPR applications in linguistics and natural language processing are introduced and some future research directions are proposed in this paper.
Key wordsStructure preference relation, ambiguous structure, knowledge acquisition, corpus.
ṹʶʹȻҪо.Ӣṹ:ʶ(Prepositional Phrase Attachment),оԱ˲ͬ緽̽,:HindleRooth1ôԶϿлȡĽĶԼʵĲͬǿų.R.Basili˽һϢ,2.CollinsBrooks3Backing-OffԴ˴.ץסӢ䷨еصѵ,ȡ°빦Ч.
ںľ䷨ṹ,о,һṹʵ:ҧ˵Ĺ4.ڴ֮,ѧַ˴Ȥṹʵ.ƹӪ5ݶЩоɹ˸ܽ,γһĺ.Щоϻڽṹöٺʵ,ȱȽϿ͹۵Ķֲ,:ʵ,ṹֵƵжṹΪĳضṹ,ϳֵĿжȵ.,ϢоԱҪΪܵṹѰЧ緽,ʵЧ±.
ľϣԺṹĶһЩ̽.ṹĶϢЧؼṹȹϵSPR(structure preference relation),Ȼͨʵıӵľ䷨ȫ,пܵSPRʵ.Զ䷨;䷨ʲԶѵ㷨ĿɹЧӦ,ʹôԭʼԶȡSPRϢΪ.Ŀǰ,ʵַ֤ĿԺʵ.
1 ṹȹϵĻ
1. Ǽдɵļ,ΪPOST.
2. ䷨Ǽо䷨ɵļ,ΪSynT.
3. ҲǾ䷨бľ䷨ṹϵıǴ,һڹϵгڹҲ,д˶.ΪRHP,RHP{POSTSynT}*.
4. γɷһֿRHPײβͬʱֵĳɷֱ,ΪIC.RHPײ,RHPṹͳΪκ,ΪSufIC.RHPβ,RHPṹͳΪǰ,ΪPreIC.IC{POSTSynT},PreIC, SufIC{POSTSynT}*.
,д䷨:P1ABP2BC,RHP,ԵõһγɷBͽǰAκC.
5. νṹɽǰγɷֺͽκ󾳹ͬɵıǴ,ΪIS,IS={PreIC IC SufIC, ICPos},ICPosʾγɷICڽνṹISеλ.
,IS={Ԧ n,1}*һĽνṹ.
6. ʵڷ,νṹеĽγɷȺߵĽǰ϶γɵĽṹʵ;ʵеĽγɷұߵĽκ϶γɵĽṹʵ.
7. ṹȹϵ(SPR)һԪ顴IS,LF,RF,ISǽνṹ,LFRFֱʾʵʵгֵƵ.
SPRеĽνṹڷеϷʽĲͬ,԰ֳ.
(1) ϽṹSPR
SPRеĽνṹγɷеһ䷨ɷ,ͼ1е(a)(c),ͼ1(a)Ϊʵ,ͼ1(c)Ϊʵ.
(2) ɢṹSPR
SPRеĽνṹĽǰγɷֺͽκ󾳷ֱɷֻҳɷϳµľ䷨ɷ,ͼ1(b)(d),ͼ1(b)Ϊʵ,ͼ1(d)Ϊʵ.

ͼ1ṹȹϵʵķֲص
SPRЩṹصΪӷԶSPRϢṩ˷.
2 SPR㷨
,ıÿӶעȷľ䷨ṹ,SPRķ־ͱúܼ,ֻ,ÿҶӽڵϷֿܵĽνṹʵ,ټSPRνṹĺԲƵϢ.
һ÷TıǴĸڵRʼ,Զ½е.Ϊ:
TreeSearch(R):
ڵRǷҶӽڵ,
ֽϽṹSPR;
ɢṹSPR;
ÿӽڵNi,ѭTreeSearch(Ni);
,.
2.1 ϽṹSPRķ
ǷеṹPH(RP1 RP2...RPn),PHΪڵ,RPi,iʣ1,nΪӽڵɷ.RP1ҲһҶӽڵ,ṹΪRP1(RP11 RP12...RP1m),䷨ɷִ:RP11 RP12...RP1m RP2...RPnһܵĽʵͼ1(a)ʾ,RP1mΪγɷ.ͬ,RPnһҶӽڵ,ṹΪRPn(RPn1 RPn2...RPnk),䷨ɷִ:RP1 RP2...RPn-1 RPn1 RPn2...RPnkһܵĽʵͼ1(c)ʾ,RPn1Ϊγɷ.
2.2 ɢṹSPRķ
ǷеṹPH(RP1 RP2...RPn),PHΪڵ,RPi,iʣ1,nΪӽڵɷ.ڷõPHеڽӳɷAi,(Ai PH)Ƿĳڵɷֵǰ׺,ṹX(Ai PH...),Ai RP1 RP2...RPnһܵĽʵͼ1(d)ʾ,RP1Ϊγɷ;ͬ,ڷõPHеڽӳɷCi,(PH Ci)Ƿĳڵɷֵĺ׺,ṹX(...PH Ci),RP1 RP2...RPn CiһܵĽʵͼ1(b)ʾ,RPnΪγɷ.
3 SPRԶȡ㷨
һڽܵSPR㷨Ҫ,ģĹҪѴ,οڻûһϴģĺ.,дԭʼı(зֺʹԱעĺı)ԶȡSPRϢо,ԵúбҪ.ԶȡSPRĹؼԭʼϿٹľ䷨ṹɭ֣Լ׼ȷؼ㲻ͬṹڷɭ֣еԤڷֲƵ.Դ,ǽ´.
3.1 ԭʼıԤ
ף6,7,һֻͳƵĺԶ䷨,ͨɷֱ߽Ԥ⡢ƥͷȽ׶εĴ,Զԭʼӵѷ.ڴ˻,ף8ݽһһָĽInside-Outside㷨ԭʼıԶѵõPCFG(probabilistic context-free grammar).Щоɹ,ǹһԭʼıԤ,ҪصǿԶԭʼӵɭ֣,Tomitaѹɭ(PSF)9ʽʾ,ѡȡеһѷ,ҷɭ֣еÿɷֽڵ϶гɷڸʺ(йڸʡԼĳɷԤڷֲƵȵϸͼ㷽ɲף8)Ϣ,˺PCFGĸϢ.
3.2 νṹķֲƵȼ
ȷķ˵,нνṹķֲƵȷ.PSF,Ͳͬ,Ϊǲ֪Ŀȷķ.,ΨһõϢһ䷨ɷAPSFеԤڷֲƵE(A).ݴ,԰·㽻νṹԤڷֲƵ.
(1) ϽṹSPRеĽνṹ.ֲƵɷֺӳɷֵԤڷֲƵȹͬȷ.,Խʵͼ1(a)ʾ,:E({PreIC IC SufIC})=E(APreIC IC).E(BA SufIC).
(2) ɢṹSPRеĽνṹ.ֲƵɽϳɷֺңɷֵԤڷֲƵȹͬȷ.,Խʵͼ1(b)ʾ,:E({PreIC IC SufIC})=E(APreIC IC).E(SufIC),E(SufIC)=E(SufIC).
4 ʵ
ʵ:(1) ӢоĲ,ģΪ1434,Լ11821,Ϊ17058,ƽ䳤Ϊ8.243/.(2) ¼СѧĿα**,ܹģΪ4139,Լ52609,Ϊ72434,ƽ䳤Ϊ12.711/.
ѡϵҪĿΪǶѱȷķṹ6,Ϊ,Ӷڴ˻ϿԷؽֲͬʽSPRȡʵ:SPRʵ飨ʵ1ԭʼıSPRԶȡʵ飨ʵ2.ͨʵıȽ,ԶĿǰSPRԶȡ㷨ĿԺʵԽбȽķ.
4.1 SPRȡʵ
5573ı,SPRȡʵ,õ½.ʵ14021ͬSPR,γSPRA;ʵ26091,ͨѡֵ=0.25,Ƶ(TF)(SPRƵ(TF)ΪĽƵ(LF)Ƶ(RF)֮,TF=LF+PF)Сڦ̵SPRų,õ3832,γSPRB.
TFֵĲͬ,ɽSPRAֳ3ӱ:(1) ƵSPR:TF10;(2) ƵSPR:1TF10;(3) ƵSPR:TF1.ȻͨӱеÿSPRSPRBǷ,ֱ¼ֵ,Լ3ӱSPRԶȡٻ.1Ǿʵ.пԿ,гֵĸƵSPR,98%϶Զ㷨ȡ,ƵSPR,ٻҲﵽ85%,ڵƵSPR,Ȼٻֻ61%ң=0.25,ʵʵʵҲϵ,ΪЩSPRһЩṹ,,PSFеԤƵȱȻǺС,,еĴ󲿷־ͻ̵ֵöų.ʹڦ=0.25,SPRԶȡ㷨ٻҲﵽ84%,ĿǰԶ㷨ȡSPRϢǿŵ.
1 SPRAĲͬӱԶȡٻ 

TF101TF10TF1ϼ
ASPR7851 8521 3844 021
B(=0.25)гֵSPR
SPRٻ(=0.25)773
0.984 71 591
0.859 1840
0.606 93 204
0.836 1
B(=0)гֵSPR
SPRٻ(=0)783
0.997 51 789
0.966 01 199
0.866 33 771
0.937 8

Ϊ˽һAͱBвͬSPRķֲ,ǴзֱѡƵߵǰ10SPR,㲻ͬSPRеĽϸLPRP,õ2Ľ.пԿ,ǻǺгĽ.㷨Ƶȼ㷽ϵĲͬ:ѷлȡƵȺʹPSFлȡԤƵ,һLFRF,LPRPķֲȴǼΪƵ:߹е8νṹ,ˡ^ n,LPRPֵǳӽ.ԶȡݺܺõطӳʵʵSPRֲ.
2 SPRƵߵǰ10SPR

SPRASPRB
νṹLFRFLPRPνṹLFRFLPRP
d^vp wD vp487.003.000.990.01m^q n469.070.001.000.00
m^q n476.000.001.000.00d^v np0.08365.960.001.00
r^vp wD vp441.003.000.990.01d^vp wD vp361.614.000.990.01
v^v n140.00292.000.320.68r^vp wD vp337.545.940.980.02
d^v vp24.00377.000.060.94v^v n147.66134.240.520.48
v^v np203.00182.000.530.47v^v np151.87128.950.540.46
d^v np3.00343.000.010.99d^v n0.00272.220.001.00
pp^v v0.00326.000.001.00r^q n272.290.001.000.00
d^v n2.00302.000.010.99d^v vp17.03246.640.060.94
d^v v3.00300.000.010.99v^np wD vp243.190.031.000.00

(ϵĽνṹʵ,'^'ʶγɷ.йصĴǺ;䷨Ǽ˵:d,wDͣٱ,m,q,n,v,r;vpʶ,npʶ,ppʶ.ݿɲף6.)
ϵķԿ,ԭʼıԶȡSPRϢ,Ǵٻ,ǴֲĸƵνṹķֲ,ܺõطӳʵʵSPRֲص.ĿǰSPRԶȡ㷨нϺõĿԺʵ.
4.2 SPRȡ
ڵʵͼش:лȡĲͬSPRĿǷϹģĲҪṩѵϲܴﵽ
ͨĿǰʵϵľȳ,γ11,ƽÿԼ507.ӵ1ʼ,ÿ1,¼ڲͬ»ȡĲͬSPR,γͼ2SPRı仯.ͼпԿ,ѵϹģĲ,ֵSPRҲڲ,ٶڲϼ,ʾ.ĿǰѵϹģ,ƲǺ.

ͼ2SPRϹģĹϵ
5  
˽ṹȹϵ(SPR)Ļ,ܶԺṹֲж,ֲͨͬ㷨ϺԭʼıԶȡ.Ŀǰʵʾ,ṹоһʵ.SPRϢ,Զֵṹ,ӴģϿԶȡͬṹʵԼھֲȵŻ.ЩʾѧоȻԴеĹӦǰ.
Ȼ,Ŀǰоֻǳ,û漰Ծ䷨Ϲϵ壨ͬһṹΪͬ䷨ɷֵ,纺еġv+nṹȿɶнṹnp:ʱ,Ҳṹvp:ӰϲϹϵ5ݵԶȡ.,ζԶȡSPRиѧ,:ЩǡǱṹ10ݣЩ塢α׼ṹ11ݣдһ̽.
л ĵоõȻѧ,ĿΪ69705005.
оõȻѧйʿѧ.
߽ܣǿ,1967,ʿ,оԱ,ҪоΪѧ.
Ʋ,1937,ڣʿʦ,ҪоΪѧ.
ͨѶϵ:ǿ 100084,廪ѧѧ뼼ϵ
עͣ*γɷڽνṹеλҴ0ʼģv--n--.
**ϵĵӰ汾ɹ¼´ѧ𶨲ʿṩڴ˱ʾл
ߵλܡǿ廪ѧѧ뼼ϵ  100084
Ʋ廪ѧܼϵͳصʵ  100084
E-mail: zhouq@s1000e.cs.tsinghua.edu.cn
ο
1Hindle D, Rooth M. Structural ambiguity and lexical relations. Computational Linguistics, 1993,19(1):103120
2Basili R, Pazienza M T, Velardi P. Semi-automatic extraction of linguistic information for syntactic disambiguation. Applied Artificial Intelligence, 1993,7:339364
3Collins M, Brooks J. Prepositional phrase attachment through a backed-off model. In: Yarowsky D, Church K eds. Proceedings of the 3rd Workshop on Very Large Corpora. Cambridge, Massachusetts: Massachusetts Insitute of Technology, 1995. 2738
4.䷨е.й,1980,(2):8192Zhu De-xi. Ambiguous phenomena of Chinese syntax. Chinese Language, 1980,(2):8190
5ݻƹӪ.ִ.о,1985,(1):6989(Huang Guo-ying. Ambiguous phrases in contemporary Chinese. Language Research, 1985,(1):6989)
6ǿ.ϿĶԶֺͱעо۲ʿѧλģ.ѧ,1996(Zhou Qiang. Phrase bracketing and annotating on Chinese language corpusPh.D. Thesis. Beijing University, 1996)
7Zhou Qiang. A statistics-based Chinese parser. In: Zhou Joe, Church K eds. Proceedings of the 5th Workshop on Very Large Corpora. Tsinghua University, Beijing, August 1997. 415
8ǿ,Ʋ.޹﷨ԶƵ.TR97002,廪ѧϵ,1997(Zhou Qiang, Huang Chang-ning. An inference approach for Chinese probabilistic context-free grammar. Technical Report TR97002, Department of Computer Science, Tsinghua University, 1997)
9Tomita M. Efficient Parsing for Natural Languagea Fast Algorithm for Practical System. Boston: Kluwer Academic Publishers, 1986
10ݷ־ΰ.ĿƼĽṹǱ.Ϣѧ,1989,3(2):115(Feng Zhi-wei. Structural description of Chinese scientific terms and potential. Journal of Chinese Information Processing, 1989,3(2):115)
11ղ.ִVPĽṹͽṹϵж˶ʿѧλģ.ѧ,1996(Zhan Wei-dong. Determining boundaries and constructional relations of verb phrase in contemporary ChineseM.S. Thesis. Beijing University, 1996)
1997-10-22յԭ,1998-02-27յ޸ĸ 
