о뷢չ
JOURNAL OF COMPUTER RESEARCH AND DEVELOPMENT
1999ꡡ3611ڡVol.36No.111999



Ͼ䷨ģʶʶĸģ
ԡƲ
ժҪȸ˺ʶ(baseNP)ʽ壬ͨȡbaseNP䷨ģ壬ʾĿɲ.ָ䷨ģֻʶbaseNPıҪǳҪ䷨ģ岢ܽbaseNPʶеı߽ģͶ.ݴˣbaseNPڲɵľ䷨ģԼNԪģͽγ˺baseNPʶģ.ʵ֤ģ͵ڵڴԱǵNԪģ.
ؼʣȻԴϿ⣬ʶ
ͼţTP391
A PROBABILISTIC CHINESE BASENP RECOGNITION MODEL
COMBINED WITH SYNTACTIC COMPOSITION TEMPLATES
ZHAO Jun and HUANG Chang-Ning
(Department of Computer Science & Technology, Tsinghua University, Beijing 100084)
(State Key Laboratory of Intelligent Technology & Management, Tsinghua University, Beijing 100084)
AbstractA formal definition of Chinese baseNP is proposed in the present paper. The manipulatability of the definition is shown through the formulation of the Chinese baseNP annotation specification oriented for information processing and the extraction of the syntactic composition templates. It is pointed out that the syntactic composition templates are only necessary but not the sufficient condition for baseNP recognition, so the boundary ambiguity and phrase-type ambiguity could not be solved depending merely on the syntactic composition templates. On the ground of this view, the basic templates embodying the baseNP composition and the N-grams modeling the contextual constraints are organically incorporated into a new probabilistic model for Chinese baseNP recognition. The experiment shows that the model is superior to the N-gram model based on part-of-speech information. 
Key wordsnatural language processing, corpus, baseNP
1
ʶ(baseNP)ʶȻԴϢͻĻо.ChurchӢbaseNPΪǶ׵ʶbaseNPʶbaseNPұ߽ǵıע⣬NԪģʵ֣1.йغbaseNPоδֻף2NԪģͽкʶʶо.ʵϢNԪģͶںʶʶǲֵ.ĽʾbaseNPڲɵĻģʾԼNԪģͽγɽϾ䷨ģʶbaseNPµĸģͣ˽ϸߵʶȷ.
2ʶĶʶ
2.1baseNPĶ
ںѧУʶĶΪ3ͣ޶ԶдԶԶ3.Ĵ޶ԶbaseNPʽ.
1. ʶ¼baseNP
baseNP  baseNP + baseNP
baseNP  baseNP + |
baseNP  ޶Զ + baseNP
baseNP  ޶Զ + |
޶Զ  ݴ|||||ִ|+ʣ
ɴˣʶԷΪbaseNP͡baseNPǻʶ¾˵.
1baseNP͡baseNPʾ

baseNP baseNP
׼Ʒṹ¸Ůӵ̨ƵĹ
̫ȻԴĴжս11ְٷչľо뷢չ

2.2baseNPʶ
baseNPĶԿbaseNPӦ÷һľ䷨ɹ򣬱ĳ֮Ϊ䷨ģ壨ģ壩.ǷϾ䷨ģֻǹbaseNPıҪǳҪ.һģĴпԲbaseNP2.
(1) ߽ģ壺ھУĳЩģĴп﷨ʽҲǷ﷨ʽ.ĽΪ߽ģ.磺N+VbaseNPһģ壬1дС족ϸģ岢baseNP2еĴС˾Ҳϸģ壬baseNPǷ﷨ʽ.
1. /N /V /V Ӫ/N ҵ/N ߳/V /N /U ·/N.
2. IBM/X ˾/N /V ȫ/A /V /N /N /U /V ۸/N.
(2) 壺ھУĳЩģĴ﷨ʽbaseNPҲ͵Ķ.ĽΪ.磺V+NbaseNPһģ壬3еĴСҵơϸģ岢baseNP4еĴСʡҲϸģ壬baseNPǶʶ.
3. /T ѧ/N ҵ/N /U ҵ/V /N Ͼ/N.
4. й/N /N /N /T /V /V /N.
ķʶbaseNP.ٴӱעbaseNPгȡbaseNPģ壬ıзЩģĴʴΪѡbaseNPԼжϺѡʴǷΪbaseNP. 
3baseNP䷨ģĳȡ
2.1ڸbaseNPʽ壬һݹ鶨.ΪʶҪȽתΪԵģ.һ2Уٽ˹עbaseNPϿ⣻ڸϿͳϢԳʼģ弯Ͻйˣģ弯.
оڳȡbaseNPģѵģΪ10֣ڶԶִʺʹԱעĻ(Աע¼)baseNP˹ע.ѵͳƵõ407baseNP䷨ģ壬гִ564Ͽ98.6%baseNPĳΪģ壬ĵоֻǻģ.2гһЩõĻģ壬ÿ䷨ɴԲַ֣֣֣.磺ģ塰NG2+VN2ʾ˫ͨʺ˫ʵϡ.
2baseNPģʾ

ģʾģʾ
B2+NG2׼A2+VN2+NG2ȷƷģ
VN2+NG2ۡ鱨NG2+NG2+VN2Ʒṹסƶȸĸ
VGO2+NG2¸ŮֶNG2+VN2+NG2鱨ʱģ

ͳʾıзģĴʴȫעΪbaseNPٻ98.6%ȷֻ48.5%.ȻֻûģʶbaseNPǲģҪôʴֵĻϢbaseNPȷ. 
4baseNPעģʽ
4.1
ھbaseNPģĳʶ֮һӿжֱעģʽĳ֮ΪbaseNPעģʽ.baseNPעģʽǴѡȷıעģʽ.磬54ֱעģʽbaseNPʶϢעģʽѡȷıעģʽģʽ4ĳ֮ΪbaseNPעģʽ.
5. ӡѧ/NG о/VNN ƶ/VGV /UTL /NG չ/VNN.
ģʽ1ѧ/NG о/VNN ƶ/VGV /UTL /NG չ/VNN.
ģʽ2ѧ/NG о/VNN ƶ/VGV /UTL /NG չ/VNN. 
ģʽ3ѧ/NG о/VNN ƶ/VGV /UTL /NG չ/VNN. 
ģʽ4ѧ/NG о/VNN ƶ/VGV /UTL /NG չ/VNN.
ӣS=w1/t1,w2/t2,,wN/wN,wiǴʣtiwiĴԱǣSNbaseNPעģʽ{M1,M2,,MN}baseNPעģʽøģ£
(1) ΪþÿһѡעģʽMָһp(M|S)
(2) ܵıעģʽM*ʹ

SbaseNPʶֻSĴԱTйأW޹.SԱʾΪt1,,tNдΪtN1,ǣbaseNPעģʽĸģͱʾ£

M={BNj1i1,,BNjkik,,BNjsisBNjkikʾtjkikעΪbaseNPbaseNPעģʽĸģΪ
(1)
4.2NԪģͷ
ף1ݡ2NԪģʵʶʶΪNԪģͿǵN֮һֹͬϵ(ĿǰʵõNԪģֻжԪԪģ)ȻҪǽṹԺͲԣΪ֮ĳϵNԪģö̾ĴֹͬϵʾĴϵǡ.磬6С/VGNŻ/NG֮ڳϵԪֻͬԪֶܱͬʾϵ.
6. /NG /A /VGN/VGO /VNO Ż/NG.
﷨ھĶĹܿĴʾ.һ˼룬еbaseNPԼΪĴʣԼֹͬϵȹԼǰʵطӳϵΪbaseNPĴڣڣֱͬbaseNPĶڣڣָͬߺԣbaseNPԼ󣬴ָ֮ͬʸ.磬62baseNPעģʽ
ģʽ1/NG /A /VGN/VGO /VNO Ż/NG.
ģʽ2/NG /A /VGN /VGO /VNO Ż/NG.
öԪ﷨һģʽ1ģʽ2ĴԱָͬʷֱΪ


ͳѧ֪ʶp(M1)p(M2)ĸֵ.
ϷĽbaseNPעģʽʱһԼƽbaseNPģNԪģУӶϺõؽbaseNPʶеĴﳤ.
4.3ڹԼNԪģͶ
Թʽ1±任

蹹baseNPĴԱǴtj1i1,,tjkik,,tjsis໥ѱעģʽļ㹫ʽΪ

ڸķսBNʾbaseNPĴԱǴֻBNйأѱעģʽļ㹫ʽ任Ϊ
(2)
ȸԼĶ壬ȻԼNԪģ.
2. Լ䣺CǾSһֱעģʽSкѡbaseNPǵĴȡõSdΪSĹԼ.
54baseNPעģʽӦĹԼֱΪ
Լ1ѧ/NG о/VNN ƶ/VGV /UTL /NG չ/VNN/.
Լ2о/VNN ƶ/VGV /UTL /NG չ/VNN/.
Լ3ѧ/NG о/VNN ƶ/VGV /UTL չ/VNN/.
Լ4о/VNN ƶ/VGV /UTL չ/VNN/.
עģʽΪMʱSĹԼΪSdӦĴԱΪt1,t2,,tL(LN)ݹԼĶ壬ɣ2ʽ任һõ¹Լѱעģʽ㹫ʽ
(3)
LǹԼĳȣsǸñעģʽеĺѡbaseNP.
3. ڸ(inner probability)֪һɷֹbaseNPóɷֵľ䷨tjiĸʳΪbaseNPڸʣʾ£
(Nji)p(tji|baseNP)
ڸʵĶ壬ѱעģʽ㹫ʽ任Ϊ
(4)
Ԫ﷨ǹԼbaseNPʶģΪ
(5)
Ӧϸģ(ʽ(5))Sעģʽ֮ǰҪԹʽеĲйƣӱעbaseNPѵйƴԱǵָͬʺģڸʣоüȻƷв.
5Ͼ䷨ģ͸ģʶbaseNP
þ䷨ģbaseNPĳʼע֮󣬿ò֪ĸģͽbaseNPעģʽ.Կһ·.
4. ·еÿһbaseNPעģʽΪһ·ÿһ·ĸʿԽƵعΪԼ·ϵдԱָͬʺģڸʵĳ˻
5. ·һӵ·о·ʵ·Ϊ·.磺ʾľ4·е·ǵ1·.
㷨. baseNPʶ㷨
룺ȷִʺʹԱעľӣ
baseNPģԾӽгʼעγɾעģʽľӣ
ڶھӵÿһ·
*Ը·baseNPעģʽԾӽйԼ
*ڹԼ䣬·ĸʣ
۽·ʵ·Ϊ·ӦbaseNPǳȡbaseNP.


ͼ1·ʾͼ
6ʵͷ
6.1׼ָ
ѵΪִʡԱע˹עbaseNP10ϣģ͵Ĳƣ Ϊ ղԼе1ϣ ŲԼѵ1. ָΪ baseNPʶľȷʺٻʣʽ£
ȷ
ٻ
aģʶȷbaseNPĸbǱģʶΪbaseNPĴʴcǲԼеbaseNP.
6.2ڴԱǵNԪģ͵ܱȽ
Ϊ˶ԵڴԵNԪģͣ2ݺͱ䷨ģ¸ģͼԶԱȣ·ֱгߵָ.пԿģ͵ľȷʺٻʾԸڵڴԵNԪģ.ӲԽӰbaseNPʶȷʵҪǱģͽ˴ľ䷨ԣûпǴʻ.
3ģ͵ܱȽ

ͻڴԵNԪģ䷨ģ¸ģ
ȷʣ%ٻʣ%ȷʣ%ٻʣ%
ղ74.278.584.893.0
Ų70.577.082.091.5

7 
һ﷨ͳƷϵĺbaseNPʶģͣģbas eNPɳɷֵĴϢϢԼbaseNPľ䷨ϢbaseNP䷨ģ ģУڵڴԱǵNԪģ.ģ͵Ĳ֮Ƕ ľ䷨ԵbaseNPʶľȷʽϵ.
*õȻѧصĿĿ69433010.
߼飺ԾУ1966ۿƼѧ²ʿооΪ˹ܡѧϢ.
ƲУ1937ڣʿʦҪоΪ˹ܡѧϢ.
ߵλ廪ѧѧ뼼ϵܼϵͳʵҡ100084
ο
1Church KA stochastic parts program and noun phrase parser fo r unrestricted text. In: Proc of the Second Conf on Applied Natura l Language Processing, Austin, Texas, 1988. 136143
2Ľݣ˺ȣϿʶԶȡΪ ԬࣺѧչӦã廪ѧ磬 1995.119125
(Li Wenjie, Pan Haihua et al. Corpus-based maximal-length Chines e noun phrase extraction. In: Chen Liwei, Yuan Qi eds. Advances and Application s on Computational Linguistics(in Chinese). Beijing: Tsinghua University Press, 1995. 119125)
3ֶ弰λйѧѧ 1996, 4 97 100
(Zhang Weiguo. Three attributes, three senses and three slots. Journal of Chinese People's University(in Chinese), 1996, 4: 97100)
4棬½. +ʡﴮǳ. йģ1996, 3: 183188
(Ma Zhen, Lu Jianming. Elementary analysis to noun + verb-type w ord stringsJournal of Chinese Language(in Chinese), 1996, 3:183188)
5۾˫V+Nṹ۷ ֣ŷ. ִ﷨о ѧ磬1995
(Shao Jingmin. Valency analysis to bi-syllables V+N structures. In: S hen Yang, Zheng Ding'ou edsResearch on Contemporary Chinese Valency Grammar(in Chinese) Beijing: Peking University Press, 1995)
ԭյڣ1997-12-01޸ĸյڣ1998-03-20.
