8_17.dvi

Transcription

1 Vol. 49 No (Aug. 2008) WFST 1, (i) (ii) 2 Weighted Finite State Transducer WFST WFST WFST-based Language Understanding for Rapid Prototyping of Spoken Dialogue Systems Yuichiro Fukubayashi, 1, 1 Kazunori Komatani, 1 Mikio Nakano, 2 Kotaro Funakoshi, 2 Hiroshi Tsujino, 2 Tetsuya Ogata 1 and Hiroshi G. Okuno 1 Language understanding (LU) modules for spoken dialogue systems in an early phase of their development need to be (i) easy to construct, and (ii) robust against various expressions. In this paper, we describe a method for constructing LU modules easily without a large amount of corpus or complicated handcrafted rules. An LU result is selected with Weighted Finite State Transducer from an automatic speech recognition output that may contain speech recognition errors. We designed several weighting schemes. A weighting scheme is determined by using training data. Since these weighting schemes are simpler than conventional methods, our method does not need a large amount of data for determining an optimal scheme. We evaluated our method in two different domains. The results revealed that our method outperformed baseline methods with less than one hundred utterances as training data, which can be reasonably prepared for new domains. This shows that our method is appropriate for a rapid prototyping of LU modules (1) (2) 1) 4) 1 Graduate School of Informatics, Kyoto University 2 Honda Research Institute Japan Co., Ltd. 1 Presently with NEC Corporation 2762 c 2008 Information Processing Society of Japan

2 2763 VoiceXML 5),6) XISL 7) 8),9) 10) VoiceXML 10) 11) 1 2) Weighted Finite State Transducer WFST 3),4) 12) 14) WFST 1 1 Fig. 1 Relationship between our method and conventional methods in spoken language understanding. 2. WFST WFST 2 (1) VoiceXML WFST (2) WFST

3 Table 1 Example utterances including FILLERs. = [month=2, day=22] Table 2 2 Language understanding results for input Every Tuesday please. Fig. 2 2 WFST FST Example of WFST for input and FST for parsing grammar. w date-repeat= 3.0 FILLER date-repeat= 2.0 FILLER FILLER date-repeat= 1.0 FILLER FILLER FILLER n/a WFST FST WFST WFST FST WFST WFST [month=2, day=22] month 2 day 22 2 WFST FST / ε ε ε ε/ FST FST FST FILLER 1 FILLER FILLER $ date-repeat= 1 WFST 2 FILLER 1 WFST WFST 2 WFST FILLER FILLER FILLER FILLER WFST FST WFST FST w 2 w 3.0 [date-repeat= ] 1 $

4 <keyphrase-class name="month">... <keyphrase> <orth> </orth> <sem>2</sem> </keyphrase>... </keyphrase-class>... <action type="specify-attribute"> <sentence> { } [*month] *day [ ] </sentence> </action> 3 Fig. 3 Example of grammar description. 3.5 WFST n-gram 3) n-gram 4) n-gram 2.2 FST 3 VoiceXML 3 keyphrase-class keyphrase 3 month 2 action sentence * keyphrase-class [] 2.1 FILLER [] {} {} FILLER WFST N -best 3 FST 3.1 N -best WFST N-best N-best N-best w i s = N j eβ score i eβ score j ws i N-best i β score i i β FST

5 2766 w w (1) word(const.): w w =1.0 (2) word(#phone): w w = l(w ) (3) word(cm): w w = CM(W ) θ w word(const.) word(#phone) W l(w ) 0 <l(w ) 1 word(#phone) word(const.) word(cm) W 15) CM(W ) θ w W W CM(W ) θ w w w W 3.3 FST w c 3 (1) cpt(const.): w c =1.0 (2) cpt(cm): W w c = (CM(W) θ c) #W (3) cpt(#p CM): W w c = (CM(W) l(w ) θ c) #W W W W #W W cpt(const.) 1 cpt(cm) 3 word(cm) cpt(#p CM) Table 3 Example of weighting for a parameter setting: word(cm) and cpt(#p CM). FILLER CM(W ) l(w ) Concept month=2 day=22 word 0.7 θ w 0.6 θ w 0.9 θ w 1.0 θ w 0.9 θ w cpt ( θ c )/1 ( θ c θ c )/2 cpt(#p CM) θ c 3.4 N-best i 1 i N 3 ws w i w w c w i i WFST i w i N-best w i = LU(argmax w i ) (1) i w i =max p i w(p i ) (2) w(p i )=ws i + α w w w + α c w c (3) W A pi W C pi LU(i) N-best i p i i WFST A pi p i C pi p i w i 3 word(cm) cpt (#p CM) α w (4.1 5θ w ) month=2 α c (0.81 θ c ) day=22 α c (0.72 θ c ) α c (1.53 2θ c ) w i ws i + α w (4.1 5θ w )+α c (1.53 2θ c )

6 2767 CM(W ) l(w ) F F month:6, day:3 F F month:6, day:3 F F F month:6, day:3, car:fit Fig. 4 F 4 Example of language understanding with WFST (rent-a-car domain). concept error rate CER CER =(S + D + I)/N ref N ref S D I C S D = N ref C S I = N hyp C S N hyp N 1 10 α w,c 1 0 θ w,c α w,c =0 4 [month:6] [day:3] [car:fit] FST 2 [month:6] [day:3] [car:fit] word(cm) 0.6 cpt(#p CM) 0.4 w = ws i w = ws i [month:6] [day:3] [car: ] 3.5 WFST FST 5 FST Fig. 5 Example of FST of a domain grammar. 6 WFST Fig. 6 Example of WFST for an input. 5 FST 6 WFST FST 2 FST *back[j] ε *back[j]

7 ,186 3, Julius N-best WFST Fig. 7 Example of WFST for an input with a weight for ASR N-bset output. j FST *back[j] 5 *back[1] ε WFST α w w w ε *back[j] /α c w c α c w c j WFST j *back[1] *back[j] ε 6 3 WFST 3 α w w w = α w (0.3 θ w ) ε *back[2] / α c (0.72 θ c ) N-best 7 N WFST ws i ε 10,000 40, % 65.7% WFST MIT FST toolkit 16) 4.2 WFST CER CER 4-fold cross validation

8 concepterrorrate CER Table 4 Concept error rates (CERs) in each domain WFST CER FST % 43.2% 4 CER WFST 5 Table 5 The optimal parameters in each domain. N α w w w α c w c word(const.) word(cm) cpt(#p CM) CER Table 6 CER when the optimal parameters are exchanged. /CER N α w w w α c w c word(const.) word(cm) cpt(#p CM) CER WFST 5 α c =0 N-best N-best CER 6 CER

9 Fig. 8 Accuracy when training data increased (video domain). 9 Fig. 9 Accuracy when training data increased (rent-a-car domain). Accuracy = 100 CER 4-fold cross validation WFST (1) WFST (2)

10 2771 WFST S B COE SCAT 1) Schatzmann, J., Weilhammer, K., Stuttle, M.N. and Young, S.: A survey of statistical user simulation techniques for reinforcement-learning of dialogue management strategies, The Knowledge Engineering Review, Vol.21, No.2, pp (2006). 2) Sudoh, K. and Tsukada, H.: Tightly integrated spoken language understanding using word-to-concept translation, Proc. EUROSPEECH, pp (2005). 3) Potamianos, A. and Kuo, H.-K.J.: Statistical recursive finite state machine parsing for speech understanding, Proc. ICSLP, Vol.3, pp (2000). 4) Wutiwiwatchai, C. and Furui, S.: Hybrid statistical and structural semantic modeling for Thai multi-stage spoken language understanding, Proc. HLT-NAACL Workshop on Spoken Language Understanding for Conversational Systems and Higher Level Linguistic Information for Speech Processing, pp.2 9 (2004). 5) VoiceXML Forum: VoiceXML Forum. 6) VoiceXML Vol.44, No.10, pp (2003). 7) XISL 2001-SLP-37-8 (2001). 8) Misu, T. and Kawahara, T.: A bootstrapping approach for developing language model of new spoken dialogue systems by selecting web texts, Proc. ICSLP, pp.9 13 (2006). 9) Weilhammer, K., Stuttle, M.N. and Young, S.: Bootstrapping language models for dialogue systems, Proc. ICSLP, pp (2006). 10) Vol.43, No.3, pp (2002). 11) Seneff, S.: TINA: A natural language system for spoken language applications, Computational Linguistics, Vol.18, No.1, pp (1992). 12) Seto, S., Kanazawa, H., Shinchi, H. and Takebayashi, Y.: Spontaneous speech dialogue system TOSBURG II and its evaluation, Speech Communication, Vol.15, No.3-4, pp (1994). 13) Kawahara, T., Lee, C.-H. and Juang, B.-H.: Flexible speech understanding based on combined key-phrase detection and verification, Speech and Audio Processing, IEEE Trans., Vol.6, No.6, pp (1998). 14) Vol.44, No.5, pp (2003). 15) Lee, A., Shikano, K. and Kawahara, T.: Real-time word confidence scoring using local posterior probabilities on tree trellis search, Proc. ICASSP, Vol.1, pp (2004). 16) Hetherington, L.: The MIT finite-state transducer toolkit for speech and language processing, Proc. ICSLP, pp (2004). ( ) ( )

11 FIT2002 ACL IEEE SFN INNS MIT 2004 ACL ACM IEEE IEA/AIE-2005 RSJ JSME JSAI IEEE AAAI 1972 NTT JST IEA/AIE IEEE/RSJ IROS Best Paper Nomination Finalist 2 JSAI RSJ ACM IEEE