100080 e-mal:{gdxe, cqzong, xubo}@nlpr.a.ac.cn tel:(010)82614468 IF 1 1 1 IF(Ingerchange Format) [7] IF C-STAR(Consortum for speech translaton advanced research ) [8] IF 2 IF 2 IF 69835003 60175012
[6][12] [2] [5] HMM [1] [2] [5] [3] IF IF 2 3 4 5 2 IF [7] [4] IF (1) c clent a agent (2) (Speech Act) gve-nformaton Pardon (3) (Concept) + reservaton room reservaton+room (4) (Arguments) Argument Value Argument Value [4] room-spec sngle double room-spec=sngle IF IF IF IF speaker: speech act+concept*(argument*) * gve-nformaton+reservaton+room(room-spec=(sngle,quantty=1)) 3 HMM HMM IF
3500 1500 3 IF HMM IF 3.1 3 [3] [10] [11] IF -----gve-nformaton+reservaton+room(room-spec(sngle,quanlty=1)) -----gve-nformaton+reservaton+room(room-spec(double,quanlty=1)) 230 N_O_ROOMLEVEL
3.2 IF IF IF Speech Act Concepts Arguments IF (1) IF Concept Argument who,room-spec (2) IF Concept Argument famly-name=,room-spec= (3) IF Concept Argument IF Concept Argument +trp:destnaton=here prce:quantty 3.3 HMM HMM HMM HMM HMM HMM HMM 1500 171 190 GREET P_PEOPLE advother V_Q_AVAILABILITY N_C_ROOM HMM greetng who nul +avalablty +room <S,O,A,B, >, [1] (1) S, N (2) O( ), M. (3) A= a j., a = P( q = S q = S ), 1, j N j t j t 1 n aj 0, aj = 1 j= 1 (4) s j v B = b ( k ), k b ( k) = P( o = v q = S ), 1 j N, 1 k M j t k t j j
m bj ( k) 0, bj ( k) = 1 k = 1 (5) π = π, π = Pq ( = S), 1 N π 0, 1 n = 1 π = 1 3.4 IF 3.4.1 HMM IF IF IF 60 IF IF HMM greetng=begn, avalablty=queston, room. IF greet=? greetng(greetng=?) (1)?=queston request-nformaton+? (2) avalablty, room avalablty+room (3) greetng=begn (1) greetng(greetng=begn) greetng=begn, avalablty=queston (2), request-nformaton+avalablty avalablty=queston avalablty,room (3) avalablty+room.avalablty,room IF greetng(greetng=begn), request-nformaton+avalablty+room IF. quantty=1, tme-unt=day who=i, dsposton=dsre room-spec= double, or, room-spec=sngle prce:quantty=100, currency=$ IF duraton=(quantty=1,tme-unt= day) dsposton=(desre,who=i) room-spec= (operator=dsjunct,[double,sngle]) prce=(quantty=234),currency= us_dollar) 3.4.2 (Speech Act) 1 IF IF
1 2 IF Concept Argument 4 1500 200 2 91.2 [2] IF 79.2 [2] [5] 3 HMM 2 [2] 91.4% 52.8% [5] 72.0% 91.2% 79.2% ( ) [2] 6439 330 737 [5] 1037 89 100 1500 171 190 3 HMM (1) (2) IF
(3) IF IF 20.8% (1) (2) Bgrams HMM Bgrams (3)IF IF 60 IF 5 IF 1 2 HMM [1],.. 1998 [2] W.Mnker, S.Bennacef. A stochastc Case Approach for Natural Language Understandng. Proc. ICSLP,1996 [3] B.Bruce. Case Systems for Natural Language. Artfcal Intellgence. 1975. Vol.6,327-360. [4] Lor Levn, Donna Gates. An Interlngua Based on Doman Actons for Machne Translaton of Task-Orented Dalogues. Proc. ICSLP,1998 [5] Yunbn.Deng, Bo Xu. Chnese Spoken Language Understandng Across Doman. Proc ICSLP, Oct, 2000. 1:230-234 [6],... 1999.115~122 [7],...1999.248~255 [8] Jun Park, Jae-Woo Yang, ETRI Speech Translaton System, C-STAR Workshop, Schwetzngen, 1999 [9]. ---... 2000.07 [10] S.K. Bennacef, H.Bonnea-Maynard. A Spoken Language System for Informaton Retreval. Proc ICSLP,1994.1271-1274 [11] M.Bates, R.Bobrow, R.Ingra. Advances n BBN s Spoken Language System. Proceedngs of the Spoken Language Technology Workshop, Mar, 1994. 43-47. [12] Chengqng Zong, Hua Wu. Analyss on Characterstcs of Chnese Spoken Language, Proc. of 5 th
Natural Language Processng Pacfc Rm Symposum, 1999. 358-362. Chnese Spoken Language Analyzng Facng the Mddle Semantc Representaton Guodong Xe, Chengqng Zong, Bo Xu Natonal Laboratory of Pattern Recognton, Insttute of Automaton Chnese Academy of Scences, Bejng, 100080 E-mal:{gdxe,cqzong,xubo}@nlpr.a.ac.cn tel:(010)82614468 Abstract: Spoken language analyzng s a crucal part n human-machne dalog system and spoken language translaton system. In ths paper we present a Chnese spoken language analyzng method based on the combnaton of statstcal and rule methods The analyzng result s a mddle semantc representaton. It has two stages, frst, use the statstcal method to analyzng the semantc nformaton, then use the rule method to map the semantc nformaton to the mddle semantc representaton. Ths method avods the shortcomng of the rule and has hgh robustness, at the same tme t acheves a lower error rate. Key words: spoken language analyzng; statstcal analyzng model; mddle semantc representaton (IF) Supported by the Natonal Natural Scence Foundaton of Chna under Grant No.69835003 and 60175012