- PowerPoint PPT Presentation

./'01$,2+34'5,.67%187, ?11/+#6%+0$*&%,(095,&"@621/"*& � , � G)=,&,L"=/*,-E& & � G/)H,4&$40,("20$)4&$-&"*-)&6-,16*&1)(&(,=)0,&2)40()*&)1&%,#$2,-&"4%& ()C)0-E& � .N'7G&I&.4&,(5)4)=$2&%$"*)5&-A-0,=&0)&2)40()*&,*,20($2"*& %,#$2,-&$4&@)6-,-&6-$45&5,-06(,&(,2)54$0$)4&& � !UQ)==,&/()D,20&I&%$"*)5-&$4&0@,&@)=,&="2@$4,&,4#$()4=,40;& � 8)54$()4&I&0@,&2)54$0$#,&()C)0&2)=/"4$)4& � 8*"($--"&Z)$2,I)/,("0,%&/()2,%6(,&C()?-,(&1)(&0@,& 340,(4"0$)4"*&G/"2,&G0"0$)4&O3GGP& & � K@,&=)-0&%,="4%$45&"//*$2"0$)4&1)(&16**A&4"06("*&"4%& 64%,(-0"4%"C*,&%$"*)5-F&"(,&#$(06"*&"5,40-&"4&2)=/"4$)4-E&& � .Z.K.T[&I&4"06("*F&$40,("20$#,&%$"*)5-&?$0@&(,-/)4-$#,&#$(06"*& @6="4-& � 8)**"5,4&I&8)4#,(-"0$)4"*&.--$-0"40-;&+6$*%$45&2)**"C)("0$#,& "5,40-&& & &

./'01$,2+34'5,.67%187, ?11/+#6%+0$*&%,(095,&"@621/"*& � , � SL"=/*,&)1&=6*0$=)%"*&%$"*)5&-A-0,=-& � G="(0H)=& & � !"#$%&"'(%'()"'*&+,(-%&'./0%,&+(.%/'121("&3'4'+&' *&+,(+5613'7%8'&+2'4')"#9'2%6:' � !)+('.1'9#+2./;'+('()"'$./"&+'(%/.;)(:' � <)"1"'+,"'()"'&%=."1'9#+2./;'(%/.;)(3'<)"'$./"&+1')+="' >""/'&+,5"?'%/'()"'&+9' '

./'01$,2+34'5,.67%187, ������������������������ ,

./'01$,2+34'5,.67%187, �� A(+"3&,+*%0(+#6/&(".+"B& � & � 34&0@,&\Z33&2,406(AF&-)=,&/@$*)-)/@,(-&*$H,& !,-2"(0,-&)(&T,$C4$J&C,5"4&?$0@&0@,&2)4-0(620$)4&)1& 64$#,(-"*&*"456"5,-&$4&1)(=&)1&46=,($2"*&2)%,-;&& & & 1 Hello Hola 2 � & � 3 � & � V$(-0&-,($)6-&"00,=/0-&0)&C6$*%&0"*H$45&-A-0,=-& ?@,(,&$4$0$"0,%&$4&0@,&\Z333&"4%&\3\&2,406($,-;& V$(-0&"60)="0"&"C*,&0)&$=$0"0,&@6="4&C,@"#$)6(& ?,(,&2)4-0(620,%;&

��������������������������� , ./'01$,2+34'5,.67%187, �� A(+"3&,+*%0(+#6/&(".+"B& � & � +"()4&Z)4&[,=/,*,4& � & ]^^_& � K"*H$45&@,"%-& � V$(-0&"60)="0"&0)&/()%62,& ?@)*,&?)(%-&"4%&-@)(0& /@("-,-;& & � Y)-,1&V"C,(& � &]`a^& � 3=$0"0$)4&)1&@6="4&-/,,2@& /()%620$)4& � S6/@)4$";&]b&H,A-&0)&1)(=& "4A&?)(%& &

./'01$,2+34'5,.67%187, ��������������������������� & � .0&0@,&,4%&)1&\3\&2,406(AF&-2$,40$-0-&O,;5;&Z)4& Q,(*=@)*0JP&2)42*6%,%&0@"0&-/,,2@&2)6*%&C,& /()%62,%&,*,20($2"**A&$4-0,"%&)1&=,2@"4$2"**A;& & This was the theory, what about the practice?? & � Yc&G0,?"(0&$4&0@,&C,5$44$45&)1&\\&2,406(A&C6$*0& "&="2@$4,&0@"0&2)6*%&5,4,("0,&#)2"*$2&-)64%-& ,*,20($2"**A;&&

./'01$,2+34'5,.67%187, ��������������������������� & � 56*' � V$(-0&,*,20($2&="2@$4,-F&1)(&,L"=/*,E& � ZB8B!SNF&-/,,2@&"4"*AJ,(&"4%&-A40@,-$J,(&%,#,*)/,%&$4&+,**& T"C)("0)($,-&0@"0&2)6*%&C,&)/,("0,%&CA&"&H,AC)"(%;& � V$(-0&0("4-*"0$45&="2@$4,-&"(,&/"0,40,%& � &V("42,&O"60)="0$2F& -$=/*,P&"4%&N6--$"&O"%#"42,%F&@6="4&$40,(#,40$)4P;& � 76*' � V$(-0&2)=/60,(-&?,(,&%,#,*)/,%&%6($45&XXd& � ]e<`& � &]ef]E&[)(4"%&g6-,F&1$(-0&/()5("=="C*,&2)=/60,(;& � ]ef<E&8)*)--6-F&+*,02@*,A&W"(H;& � ]ef<& � &]efaE&74$#;&W,44-A*#"4$"F&S>3.8;& � K@,&/)0,40$"*&)1&0@,-,&="2@$4,-&C,5$4-&0)&C,& 2)4-$%,(,%&1)(&)0@,(&/6(/)-,-E& � .*"4&K6($45& � X"((,4&X,"#,(& &

./'01$,2+34'5,.67%187, ��������������������������� & � 86*' � a_&I&.*"4&K6($45& � &340,**$5,40&9"2@$4,-& � a]& � &8))/,(F&T$C,(="4&"4%&+)(-0& � &G/,20)5("=& � ad& � &9"2@$4,&0("4-*"0$)4&2)41,(,42,-& � a^& � &8@)=-HA& � &GA40"20$2&G0(6206(,-& � a`& � &W,0,(-)4F&X"45&"4%&G$#,(-0,4& � &!$/@)4,-& � 96*' � V$(-0&2)4#,(-"0$)4"*&"5,40-F&,;5;&X,$J,4C"6=&ST3g.& � &>)0&2)=/60"0$)4"*&*$456$-0$2-& � :6*' � 8)=/60"0$)4"*&*$456$-0$2& � +"-,%&$4&0@,)($2"*&?)(H&"C)60&-,="40$2&"4%&*"456"5,& 2"(($,%&)60&CA&8@)=-HAF&9)40"56,F&X))%& � N6*,&C"-,%&-/,,2@&-A40@,-$-&

./'01$,2+34'5,.67%187, ��������������������������� & � ;6*' � K,*,/@)4A&-,(#$2,-& � !$20"0$)4&-A-0,=-& � <6*' � 8)(/)("&%,#,*)/=,40E&X)(%4,0& � 3-)*"0,%&?)(%&(,2)5;&K,*,/@)4A&%$"*)5&-A-0,=-& � !,%$2"0,%&2)=/"4$,-;& � =666*' � G0"0$-0$2"*&-A-0,=-& � &N6*,-&-0$**&$4&/("20$2"*&6-,& � G0"4%"(%-&1)(&,"-A&%,#,*)/=,40;& � =6>6*' � 7-"C$*$0AF&6-,(&2,40(,%&-A-0,=-;& &

./'01$,2+34'5,.67%187, ?(#,+%"#%9("&

./'01$,2+34'5,.67%187, ?(#,+%"#%9("& INPUT OUTPUT INT ERFA C E INT ERFA C E System User voice voice ASR NLP NLG DM SS

./'01$,2+34'5,.67%187, ?(#,+%"#%9("& Conv e r s ational Ag e nt Integration of the different agents

./'01$,2+34'5,.67%187, ?9%026%+#&41""#,&!"#05$+%+0$& � G)=,&.GN&2)=/*,L$0$,-E& � T$456$-0$2&#"($"C$*$0A& User � 340,(I-/,"H,(&#"($"C$*$0A& voice ASR NLP � 340("I-/,"H,(&#"($"C$*$0A& � 8@"44,*&#"($"C$*$0A& & & � 9"AC,&A)6&2)6*%&"%%&A)6(&)?4&0)&0@,&*$-0RR&

./'01$,2+34'5,.67%187, ?9%026%+#&41""#,&!"#05$+%+0$& � (+#+"*+"0#$'.//?%#0@2''A#?B%C'D"33,1'A%3,$*' � BCD,20$#,E&K)&1$4%&0@,&-,h6,42,&)1&?)(%-&X&5$#,4&"& -,h6,42,&)1&"2)6-0$2&%"0"&.E& & Acoustic Model & & Language Model & & � K("$4$45&%"0"i& � ����������������������������������� & & & &

./'01$,2+34'5,.67%187, C6%9(6/&D6$5965"&E(0#"**+$5& In dialog systems: Obtains the meaning of the recognized user utterance ASR NLP DM In a wider sense: Understanding human natural language It has many applications, among them: Levels of NLP: � Machine translation - Morphological � Question answering, text - Lexical summarization, text simplification � Information classification, filtering and - Sintactical recovering, e.g. spam filters, orthographic - Semantic correctors - Discourse � Dialog systems � - Pragmatical

./'01$,2+34'5,.67%187, C6%9(6/&D6$5965"&E(0#"**+$5& � K@,&5)"*&)1&-,="40$2&"4"*A-$-&$-&0)&(,/(,-,40&0@,&=,"4$45&)1& 0@,&-,40,42,-&$4&"&*)5$2"*&*"456"5,&

IJK,E,2+7*')&71, ./'01$,2+34'5,.67%187, C6%9(6/&D6$5965"&E(0#"**+$5& � 342(,=,40"*&340,(/(,0"0$)4& � K@,&%$-2)6(-,&$-&"&-,h6,42,&)1&-,40,42,-&0@"0&2"4&C,&-0(6206(,%&$4&"& ?"A&)(&"4)0@,(&0)&/()%62,&%$11,(,40&=,"4$45-;& � N@,0)($2& � N,1,(,42,&(,-)*60$)4& � .4"/@)("& � QA/)4A=-& � W()4)64-& � >>-&O/;,;&X@"0&?"-&/(,#$)6-*A&2)==,40,%P;& � B0@,(&,L/(,--$)4-&O/;,;&@,&?"-&1,,%$45&0@,&/$5-P;& � X)(*%&H4)?*,%5,&4,,%,%& � G6//)-$0$)4-&O,;5;&3&2"=,&$4&=A&C$H,;&K@,&?@,,*&="H,-&"&-0("45,& 4)$-,P;& � ���������������������������������������������������� &

./'01$,2+34'5,.67%187, C6%9(6/&D6$5965"&E(0#"**+$5& IJK,E,=8?+5)+%6, � E%-/$,F"+)'%G'1#+4?#$'$#1&4#&,'"*'"1'"+*'#-H"&4"+)I'J#+4?#$' $#1&4#&,'"*'#-H"&4%4*'#+'C#?"%4*'$,C,$* E& & � T,L$2"*&& � G"=,&?)(%&=$5@0&@"#,&%$11,(,40&5("=="0$2"*&2"0,5)($,-;& � GA40"20$2& � W@("-,-&?$0@&=)(,&0@"4&)4,&"--)2$"0,%&-A40"20$2&-0(6206(,;& � G,="40$2& � G"=,&?)(%&%$11,(,40&=,"4$45-;& � G"=,&-A40"20$2&-0(6206(,F&%$11,(,40&=,"4$45-;& � !$-2)6(-,& � N,1,(,40$"*& � W("5="0$2& � ����������������������������� &

./'01$,2+34'5,.67%187, C6%9(6/&D6$5965"&E(0#"**+$5& � B("*&4"06("*&*"456"5,&@"-&/,26*$"($0$,-&0@"0& ="H,&$0&,#,4&=)(,&%$11$26*0&0)&/()2,--E& � TS\38.T& � G,#,("*&6-,(-&0"*H$45&"0&-"=,&0$=,& � &G,/"("0$)4& � Q)=)/@)4$2&?)(%-& � 8)**)h6$"*&1)(=&)1&?)(%-& � V("5=,40-F&@,-$0"0$)4-F&(,/,0$0$)4-& � c6"-$*,L$2"*&O,;5;&"@"P& � 9BNWQBTB'38.T& � Q,-$0"0$)4-& � X)(%&2*"--,-&0@"0&)4*A&"//,"(&$4&)("*&*"456"5,F&*$H,&1)(& ,L"=/*,&%$-2)6(-,&="(H,(-&O,;5;&?,**P& � 340,(D,20$)4-& � 96*0$I?)(%&,L/(,--$)4-&O,;5;&3&-,,P& &

./'01$,2+34'5,.67%187, C6%9(6/&D6$5965"&E(0#"**+$5& � GA40"20$2& � 342)=/*,0,&-0(6206(,-& � @,"'2%6'%5:' � ���������������������������� ' � !%,,."?:' � A"1' � N,/"("0$)4-& � B/'C%/?+2D'%/'<6"1?+2D'4'8"/('(%'()"''>"+$)' � N,/,0$0$)4-& � C2'1.1(",D'&2'1.1(",'.1'=",2'0,."/?#2' � GA40"20$2&2)=C$4"0$)4-& � ����������������������������������������������� ' '

./'01$,2+34'5,.67%187, C6%9(6/&D6$5965"&E(0#"**+$5& � K%**"H$,'"-/$,-,1+#+"%1*' � L4$,MH#*,3'-,+@%3%$%&",*I' � SL0("20&0@,&-,="40$2&$41)(="0$)4&1()=&"&-A40"LI-,="40$2& "4"*A-$-&)1&0@,&-,40,42,-E& � G/,2$1$2&5("=="(-&%,1$4,%&1)(&0@,&0"-H;&& � !,0,20$)4&)1&-$45*,M-,h6,42,-&)1&H,A&?)(%-&?$0@& -,="40$2&=,"4$45;&& � (+#+"*+"0#$'-,+@%3%$%&",*I' � !,1$4$0$)4&)1&*$456$-0$2&64$0-&?$0@&-,="40$2&2)40,40&"4%& 2)4-0(620$)4&)1&"&-0"0$-0$2"*&=)%,*&1()=&"&2)(/6-&)1&*"C,*,%& -"=/*,-;& � K("$4$45&0@,&=)%,*E&&8"/06(,&2)((,-/)4%,42,-&C,0?,,4&$4/60& 0,L0&"4%&$0-&2)((,-/)4%$45&-,="40$2&$40,(/(,0"0$)4;& � K("$4$45&%"0"ii&

./'01$,2+34'5,.67%187, C6%9(6/&D6$5965"&E(0#"**+$5& B2BCLI,/&'M1*%,, Semantics of the task: USERS: Semantic interpretation using frames (concepts and attributes). � Task-dependent concepts: Availability , Booking , Booked and Cancellation. � Task-independent concepts: Acceptance , Rejection and Not-Understood . � Attribute: Sport , Hour , Date , Court-Type , Court-Number and Order- � Number . Yes, I want to book a tennis court for tomorrow afternoon. (Affirmation) (Booking) Sport: tennis Date: tomorrow Hour:afternoon

./'01$,2+34'5,.67%187, C6%9(6/&D6$5965"&E(0#"**+$5& B2BCLI,/&'M1*%,, Semantics of the task: SYSTEM: Semantic representation using frames. � Concepts: Availability, Booking, Booked, Cancellation , Sport, Date, Hour, � Court-Type, Confirmation-Availability, Confirmation-Booking, Confirmation- Booked, Confirmation-Cancellation, Confirmation-Sport, Confirmation-Date, Confirmation-Hour, Confirmation-CourtType, Rule-Info, Booking-Choice. Atributtes: Sport , Hour , Date , Court-Type , Court-Number , Order-Number y � Availability-Number . Do you want to book the squash court number 1 for February 2nd? (Confirmation-Booking) Sport: squash Date: 02-02-2012 Court-Number:1

./'01$,2+34'5,.67%187, C6%9(6/&D6$5965"&E(0#"**+$5& slots 1 Process of unification of 2 frames: to combine 3 4 several frames to 5 frames generate a new frame 6 7 8 . . . . . . n

./'01$,2+34'5,.67%187, (Viajero system) S: Welcome to the system... # FR STAT ORG DEST TYPE HOUR 1 U: I want to go to Madrid Madrid S Tell me the origin city U: From Valencia 2 Valencia Unification 3 Valencia Madrid frames 1 and 2 S: Do you want a direct bus? 4 U: Yes Direct Unification 5 Valencia Madrid Direct frames 3and 4 S: There are four buses from Valencia to Madrid at 9, 10, 11and12. Tell me the departure hour 6 U: At 12 12:002 Unification 7 Valencia Madrid Direct 12 <a& 12:002 frames 5 and 6 35

./'01$,2+34'5,.67%187, -+6/05&26$65"2"$%& NLP NLG DM Inteligence � Decision - Coordination - Deals with different sources of information - NLP results, database query results - Application domain � World knowledge complex - Knowledge about users and their intentions - Dialog history - Depends on: - Task - Flexibility - Initiative

./'01$,2+34'5,.67%187, -+6/05&26$65"2"$%& � K@,(,&$-&4)0&"4&"5(,,=,40&"C)60&?@"0&0@,&0"-H&)1&"& %$"*)5&="4"5,(&"(,;&Q,(,&0@,(,&$-&"&-@)(0&*$-0E& � !,2$%,&0@,&4,L0&-A-0,=&06(4&"4%&?@$2@&$41)(="0$)4&0)&/()#$%,;& � X@"0&$41)(="0$)4&0)&/()#$%,R&Q)?R&& � X@,(,&0)&5,0&0@,&$41)(="0$)4R&9"4"5,&=6*0$/*,&-)6(2,-&& � Q)?&$-&0@,&%,2$-$)4&0"H,4R&V()=&*"-0&6-,(&06(4R&V()=&?@)*,&%$"*)5& @$-0)(AR& � 9"4"5,&$40,("20$)4& � X@)&@"-&0@,&$4$0$"0$#,&$4&0@,&%$"*)5R&& � 9"4"5,&,(()(-&$4&(,2)54$0$)4&)(&64%,(-0"4%$45&=)%6*,-& � 3-&0@,&$4/60&0)&0@,&="4"5,(&2)((,20R& � 31&4)0F&@)?&0)&2)((,20&$0R& � 9"4"5,&2)41$(="0$)4-& � X@$2@&/$,2,-&)1&$41)(="0$)4&"(,&4,2,--"(A&0)&2)41$(=R& � Q)?R&& � [,,/&0("2H&)1&-A-0,=&@$-0)(A& � X@$2@&$41)(="0$)4&$-&4,2,--"(A&0)&-0)(,R& � Q)?R& &

./'01$,2+34'5,.67%187, -+6/05&26$65"2"$%& � E%10,/+*'"1C%$C,3 E& & � N1"+"#+"C, &I&K@,&$40,(*)260)(&?@)&2)40()*-&0@,&%$"*)5& "0&"&2,(0"$4&/)$40&@"-&0@,& !"!#!$#!%&' "0&0@$-&/)$40&"4%& %,2$%,-&)4&?@"0&0)&0"*H&"C)60&4,L0;& � E%1G"?-#+"%1 &I&GA-0,=&?"A&)1&2@,2H$45&?@,0@,(&0@,& *"-0&/$,2,&)1&$41)(="0$)4&5"0@,(,%&1()=&0@,&6-,(&$-& 2)((,20&)(&4)0;&& � K?,3"0+"%1 &I&SL/,20"0$)4&)4&?@"0&0@,&6-,(&?$**&-"A& 4,L0;& � &7-,(&=)%,**$45;& � !"#$%&'@"*+%?)' I&T)5&)1&$41)(="0$)4&0@"0&@"-&"*(,"%A& C,,4&,L2@"45,%&0@()65@)60&0@,&%$"*)5;&30&2"4&C,&-"#,%& $4&1)(=&)1&%$"*)5&"20-&*$-0;&

./'01$,2+34'5,.67%187, -+6/05&26$65"2"$%& � 34$0$"0$#,&2"4&C,E& � GA-0,=&%$(,20,%& � 7-,(&%$(,20,%&O/$)4,,(E&Q)?&="A&3&@,*/&A)6R&.KjKP& � 9$L,%& & � G0"0$2& � &.*?"A-&0@,&-"=,&$4$0$"0$#,&0A/,& � !A4"=$2& � &34$0$"0$#,&2@"45,-&%6($45&0@,&$40,("20$)4& & � 8)41$(="0$)4&-0("0,5$,-E& � 3=/*$2$0& Control Flexibility � SL/*$2$0& &

./'01$,2+34'5,.67%187, C6%9(6/&/6$5965"&5"$"(6%+0$& Generates natural language from a machine representation of the NLG DM SS content to be conveyed Id-message: 000 Relation: SHOWING Arguments: 5 basic steps: Film: VÉRTIGO Room: 2 � Content organization � Discourse Session: 18:30 planning � Content distribution in sentences � Lexicalization � Referential expressions generation � Linguistic realization

./'01$,2+34'5,.67%187, C6%9(6/&/6$5965"&5"$"(6%+0$& � !$-2)6(-,&/*"44$45E& � G0(6206(,&"4%&)(5"4$J,&=,--"5,-&0)&C,&5,4,("0,%;& � 30&2"4&@"#,&-,#,("*&*,#,*-&)1&2)=/*,L$0AE& � B(5"4$J"0$)4&C"-,%&)4&/*"44$45& � &.(0$1$2$"*& $40,**$5,42,F&/(,2)4%$0$)4-&"4%&,11,20-;& � B(5"4$J"0$)4&C"-,%&)4&0,=/*"0,-& � &G="**&46=C,(&)1& 2@)$2,-&0@"0&2"4&C,&5,4,("0,%&$4-0"40$"0$45& 0,=/*"0,-;& ' ' ' � 8)4D6420$)4& � S**$/-$-& � +6$*%$45&)1&?)(%&-,0-& � 8)=/*,L&-,40,42,-&

./'01$,2+34'5,.67%187, C6%9(6/&/6$5965"&5"$"(6%+0$& Text with numbers, symbols, etc. � K("4-1)(=-&0@,&$4/60&0,L0&$40)& "&-A=C)*$2&(,/(,-,40"0$)4;& ]; K("4-1)(=-&46=C,(-F& Text pre- "CC(,#$"0$)4-&"4%&-A=C)*-&$40)& processing or normalization ?)(%-;&& d; .--$54-&/@)4,0$2& Text 0("4-2($/0$)4-&0)&,"2@&?)(%& composed "4%&%$-0$456$-@,-&/()-)%$2& only by words 64$0-&*$H,&/@("-,-&"4%& -,40,42,-;&& X-to-phoneme '("/@,=,I0)I/@)4,=,& � & � 'KW& K,L0I0)I/@)4,=,&&I&KKW& � Symbolic linguistic representation

./'01$,2+34'5,.67%187, C6%9(6/&/6$5965"&5"$"(6%+0$& � !$11,(,40&0,2@4$h6,-&)1&G/,,2@&GA40@,-$-E ' � +"-,%&$4&@6="4&-/,,2@&/()%620$)4& � W"("=,0($2& � &9)%,*&-,#,("*&/@A-$)*)5$2"*& /"("=,0,(-&"C)60&#)2"*&0("20&0)&/()%62,&-)64%-;& � +"-,%&$4&1)(="40-& � &V)(="40-&=)%,*&#$C("0$)4&)1& #)2"*&2@)(%-;&& � X$0@)60&/@A-$)*)5$2"*&C"-,& � 8)42"0,4"0$#,&-A40@,-$-& � &Y)$4&/(,I(,2)(%,%&64$0-& )1&@6="4&#)$2,& � &74$0&-$J,RR& Naturalness Automatization

89%/+$"& !" #$%&'()*%+'$, -" ./'01$,2+34'5,.67%187, F; !"#"$%&'("$)*G&?)61%6%+0$& 3> .%3%+7%+*34,(+34'5,83$35181$%, ?> @71&,A'(14+$5, *> B8'%+'$,:1*'5$+%+'$, (> C'$%1D%E3F3&1,767%18, 1> A)4%+8'(34,#$%1&3*%+'$, G" =//4+*3%+'$,('83+$7, H" C'$*4)7+'$7,

&

:1*1$%,;&1$(7, !"#"$%&'("$)*& � W()"20$#$0A& � .=3&,4#$()4=,40-& � .60)="0$2&(,-)*60$)4&)1&/()C*,=-& � B/0$="*&%$"*)5&-0("0,5$,-& � 8)40,L0I"?"(,&-A-0,=-& � S=)0$)4"*&$40,**$5,42,& � W)(0"C$*$0A& ADAPTATION � 96*0$=)%"*F& =6*0$*"456"5,F=6*0$%)="$4& -A-0,=-& � N,2)==,4%$45&-A-0,=-& � 8)=/"("0$#,&,#"*6"0$)4& � K$=,&"4%&,11)(0&=$4$=$J"0$)4& &

89%/+$"& !" #$%&'()*%+'$, -" ./'01$,2+34'5,.67%187, F; !"#"$%&'("$)*G&?)61%6%+0$& 6H 4%6%+*%+#6/&)+6/05&26$65"2"$%& ?> @71&,A'(14+$5, *> B8'%+'$,:1*'5$+%+'$, (> C'$%1D%E3F3&1,767%18, 1> A)4%+8'(34,#$%1&3*%+'$, G" =//4+*3%+'$,('83+$7, H" C'$*4)7+'$7,

:1*1$%,;&1$(7, -+6/05&I6$65"2"$%& � !,-$54&)1&%$"*)5&-A-0,=E&& � G/,2$1$2"0$)4&)1&&0@,&-A-0,=&"20$)4-&$4&(,-/)4-,&0)&6-,(&600,("42,-& "4%&,4#$()4=,40"*&-0"0,-&& � !$"*)5&-0("0,5A&0@"0&%,1$4,-&0@,&2)4#,(-"0$)4"*&C,@"#$)(&)1&0@,& -A-0,=;&& & & � '(,"0&,11)(0&$-&,=/*)A,%&0)&,=/$($2"**A&%,-$54&%$"*)5&-0("0,5$,-&1)(& 2)==,(2$"*&-A-0,=-&O4)0&"&0($#$"*&0"-HP&

:1*1$%,;&1$(7, -+6/05&I6$65"2"$%& � O%"0,PAQ &O OPAQR' � X<8U-&-0"4%"(%&\9T&1)(="0&1)(&-/,2$1A$45&$40,("20$#,&#)$2,& %$"*)5-&C,0?,,4&"&@6="4&"4%&"&2)=/60,(;&& � 30&"**)?-&#)$2,&"//*$2"0$)4-&0)&C,&%,#,*)/,%&"4%&%,/*)A,%&$4&"4& "4"*)5)6-&?"A&0)&QK9T&1)(&#$-6"*&"//*$2"0$)4-;&& � Y6-0&"-&QK9T&%)26=,40-&"(,&$40,(/(,0,%&CA&"&#$-6"*&?,C& C()?-,(F&Z)$2,\9T&%)26=,40-&"(,&$40,(/(,0,%&CA&"&#)$2,& C()?-,(;& � S#)*60$)4E& � C>I6'S=666R' � C=I6'S=667R' � C=I>'S=66:R' http://www.w3.org/TR/voicexml20/ � O5I6'S=6>6R' http://www.w3.org/TR/voicexml21/

:1*1$%,;&1$(7, Recommendation Proposed Recommendation Candidate Semantic Recommendation Interpret- Voice XML 2.1 Ration Voice Grammar Synthesis Last Call (SISR) XML 2.0 (SRGS) (SSML) Working Draft Call Control (CCXML) Working Draft Voice XML 3.0 Requirements

:1*1$%,;&1$(7, -+6/05&I6$65"2"$%& E'S0%-/4+,?R2'T,$0%-,'+%'+@,'U,#+@,?' "1G%?-#+"%1'*,?C"0,I'T@#+'*+#+,V' D'S@4-#1R2'D,$/' E2'K$,#*,'*/,#B'+@,'*+#+,'G%?'U@"0@')%4'U#1+' +@,'U,#+@,?I' D2'W,%?&"#' E2'T@#+'0"+)V' D2'XH$"*"' E2'N'3"3'1%+'413,?*+#13'U@#+')%4'*#"3I'T@#+' 0"+)V' D2'A#0%1' E2'X@,'0%13"+"%1*'"1'A#0%1'W,%?&"#'#?,' *411)'#13'0$,#?'#+'>>'.A'III' '

:1*1$%,;&1$(7, ROOT app- root.vxml A P P L I C A T LEAF I leaf.vxml O N

:1*1$%,;&1$(7, -+6/05&I6$65"2"$%& 8E&X,*2)=,&@)=,;&G"A&)4,&)1E&-/)(0-k&?,"0@,(k&G0"(5"J,(&"-0()/@A-$2-&4,?-;& QE&.-0()*)5A;& 8E&3&%$%&4)0&64%,(-0"4%&?@"0&A)6&-"$%;&O"&/*"01)(=I-/,2$1$2&%,1"6*0&=,--"5,;P&& 8E&X,*2)=,&@)=,;&G"A&)4,&)1E&-/)(0-k&?,"0@,(k&G0"(5"J,(&"-0()/@A-$2-&4,?-;& QE&-/)(0-;& 8E&O/()2,,%-&0)&@00/EMM???;-/)(0-;,L"=/*,;2)=M#L=*M-0"(0;#L=*P&& &

:1*1$%,;&1$(7, -+6/05&I6$65"2"$%& GRAMMAR SRGS - XML

:1*1$%,;&1$(7, -+6/05&I6$65"2"$%& GRAMMAR SRGS - ABNF

:1*1$%,;&1$(7, -+6/05&I6$65"2"$%& Z)$2,\9T&"(2@$0,206(,&1)(&0@,&%,-$54&)1&%$"*)5&-A-0,=-&

:1*1$%,;&1$(7, -+6/05&I6$65"2"$%& & � ZB\SB&8B997>3Kl&O@00/EMM2)==64$0A;#)L,);2)=P& & � D,).1"+#'L#/"3.//'S@++/2YYG?,,*/,,0@I@,)#1"+#I0%-R' � X,$$-,'!,H4&'Q%&'S@++/2YY*+43"%I+,$$-,I0%-R' � O%"0,'W,1",'E#$$'Q%&'ZF/$%?,?'S@++/2YY3,C,$%/,?IC%"0,&,1",I0%-R' � [,O%0#$'O%0#$'(0?"/+,?'S@++/2YY0#G,IH,C%0#$I0%-R' � .//$,'S@++/2YYUUUI#//$,I0%-Y-#0%*Y*/,,0@R' � .X\X'S@++/2YYUUUI#++I0%-Y#*/&R' � [[J'S@++/2YYUUUIHH1I0%-Y3,/#?+-,1+*Y3,/+M*$/I@+-$R' � N[A'S@++/2YY"H-I0%-Y*%GU#?,Y*/,,0@R' � Q,?1%4+'\'D#4*/",'S@++/2YYUUUI$@*I0%-R' � A"0?%*%G+'S@++/2YYUUUI-"0?%*%G+Y*/,,0@R' � J4#10,'S@++/2YYUUUI14#10,I0%-R' � K@"$"/*'S@++/2YYUUUI*/,,0@IH,I/@"$"/*I0%-R' � (/,,0@U%?B*'S@++/2YYUUUI*/,,0@U%?B*I0%-R' ' '

:1*1$%,;&1$(7, -+6/05&I6$65"2"$%& � B42,&0@,&%$"*)5&-0("0,5A&@"-&C,,4&%,-$54,%F&0@,&$=/*,=,40"0$)4&)1& 0@,&-A-0,=&2"4&C,&2"(($,%&)60&6-$45&0@,&-0"4%"(%&Z)$2,\9T ������ @)?&0)&%,-$54&0@$-&-0("0,5AR&& � .//*$2"0$)4&)1&-0"0$-0$2"*&"//()"2@,-&0)&%$"*)5&="4"5,=,40& & & � B7N&GBT7K3B>E&&8)=C$4,&0@,&1*,L$C$*$0A&)1&-0"0$-0$2"*&%$"*)5& ="4"5,=,40&?$0@&0@,&1"2$*$0$,-&0@"0&Z)$2,\9T&)11,(-F&& � 340()%62,&-0"0$-0$2"*&=,0@)%)*)5$,-&1)(&0@,&%,#,*)/=,40&)1& 2)==,(2$"*&O"4%&4)0&-0($20*A&"2"%,=$2P&2)4#,(-"0$)4"*&"5,40-;&

:1*1$%,;&1$(7, 4%6%+*%+#6/&-+6/05&I6$65"2"$%& � K@,&6-,&)1&-0"0$-0$2"*&=)%,*-&0@"0&"(,&"60)="0$2"**A&*,"(40&1()=&%"0"& @"-&/()#$%,%&#,(A&$40,(,-0$45&(,-6*0-E& � G/,,2@&(,2)54$0$)4;& � T"456"5,&64%,(-0"4%$45;& � !"#$%&'-#1#&,-,1+ ;& & � K@,&/()C*,=&)1&2)#,("5,E& � .**&0@,&/)--$C*,&-$06"0$)4-&0@"0&2"4&)226(&%6($45&"&%$"*)5&=6-0&C,& =)%,*,%& 6-$45& )4*A& 0@,& 0("$4$45& 2)(/6-& � & E$#**"G"0#+"%1' /?%H$,- ;& � 8)4-$%,("C*,& ,11)(0& $-& 4,,%,%& 0)& "2h6$(,& "4%& *"C,*& "& %$"*)5& 2)(/6-& � & .4+%-#+"0'3"#$%&'&,1,?#+"%1 ;&

:1*1$%,;&1$(7, 4%6%+*%+#6/&-+6/05&I6$65"2"$%& WB9!W&(,/(,-,40"0$)4E& �

:1*1$%,;&1$(7, 4%6%+*%+#6/&-+6/05&I6$65"2"$%& WB9!W&(,/(,-,40"0$)4E& � & +,*$,1&-/"2,E& � & & N,?"(%E& � & & & & Using Point-based Value Iteration or Q-learning, in combination � with a user simulator: & &

:1*1$%,;&1$(7, 4%6%+*%+#6/&-+6/05&I6$65"2"$%& Most slot-filling dialog systems: the dialog manager takes its decisions based � only on the information provided by the user in the previous turns and its own model (DIHANA task). Formal description: � � A i : the system answer at time i . � U i : semantic representation of the user turn at time i . Dialog � S i : State of the dialog sequence at time i

:1*1$%,;&1$(7, 4%6%+*%+#6/&-+6/05&I6$65"2"$%& Formal description: � At time i , the objective of the dialog manager is to find the best system � answer A i The selection is made by maximizing: � All the possible system answers. We establish a partition in the space of sequences of states: � DR i : Dialog register at time i (concepts and attributes). -

:1*1$%,;&1$(7, 4%6%+*%+#6/&-+6/05&I6$65"2"$%& Formal description: � For a sequence of states of a dialog, there is a corresponding sequence � of DR : Two different sequences of states are considered to be equivalent if they � lead to the same DR i Great reduction in the number of different histories in the dialogs. � A loss in the chronological information. �

:1*1$%,;&1$(7, 4%6%+*%+#6/&-+6/05&I6$65"2"$%& Formal description: � The selection of the best A i is given by: � Each user turn: � supplies the system with information about the task; � provides other kinds of information, such as task-independent � information ( Affirmation , Negation , and Not-Understood dialog acts).

:1*1$%,;&1$(7, 4%6%+*%+#6/&-+6/05&I6$65"2"$%& 2#N=I=,/&'M1*%,, Dialog Register representation: � The DR is a sequence of 15 fields: � Five concepts: Hour , Price , Train-Type , Trip-Time , and Services . � Ten attributes: Origin, Destination, Departure-Date, Arrival-Date, Departure- � Hour, Arrival-Hour, Class, Train-Type, Order-Number, and Services. We have assumed that the exact values of the attributes are not � significant to determine the next system answer: 0: The concept is not activated, or the value of the attribute is not � given. 1: The concept or attribute is activated with a confidence score that � is higher than a given threshold. 2:The concept or attribute is activated with a confidence score that � is lower than the given threshold. DR = 15 length string of elements from {0,1,2}. �

:1*1$%,;&1$(7, 4%6%+*%+#6/&-+6/05&I6$65"2"$%& STATISTICAL DIALOG MANAGER User Turn DR DR i-1 , S i-1 A 1i System (frames) MLP S Response

:1*1$%,;&1$(7, 4%6%+*%+#6/&-+6/05&I6$65"2"$%& Example of a dialog: � System 1 : Welcome to the railway information system. How can I help you? A 1 : Opening DR: 00000-1000001000 User 1 : I want to go to Barcelona. U 1 =() DR: 00000-1100001000 + Opening + U 1 � A 2 = (Confirmation:Departure-Hour:Nil) System 2 : Do you want to know the timetables? User 2 : Yes, for the Euromed train. U 2 =(Affirmation) DR: 10000-1100001100 + Confirmation:Departure-Hour + U 2 � A 3 : (Question:Departure-Date:Nil) System 3 : Tell me the departure date.

:1*1$%,;&1$(7, Example of a dialog: � User 3 : Tomorrow U 3 =() DR: 10000-1120001100 + Question:Departure -Date + U 3 � A 4 = Confirmation:Departure-Date System 4 : Do you want to leave tomorrow? User 4 : Yes U 4 =(Affirmation) DR: 10000-1110001100 + Confirmation:Departure-Date + U 4 � A 5 = Answer:Departure-Hour:Number-Trains,Train-Type,Departure-Hour New-Query System 5 : There are several Euromed trains. The first one leaves at 08:54 and the last one at 23:45. Anything else? User 5 : No, thank you U 5 =(Negation) DR: 10000-1110001100 + Answer-New-Query:Departure-Nil + U 5 � A 6 = Closing System6: Thanks for using this service. Have a good trip.

:1*1$%,;&1$(7, 4%6%+*%+#6/&-+6/05&I6$65"2"$%& More complex dialog systems: the dialog manager selects the system turn � taking into account: the information provided by the user. � the information generated by the module that controls the application. � Application manager ( AM ): � Performs the queries to the database. � Verifies if the user query follows the regulations defined for the task. � The result of queries to the AM has to be considered in order to generate the following system turn.

:1*1$%,;&1$(7, 4%6%+*%+#6/&-+6/05&I6$65"2"$%& We have established two phases for the selection of the next system turn in this type of tasks: First phase: Select the best request to be made to the AM. � Second phase: Generate the final system turn. � We propose the use of a multilayer perceptron (MLP) to obtain the system answer.

:1*1$%,;&1$(7, 4%6%+*%+#6/&-+6/05&I6$65"2"$%& STATISTICAL DIALOG MANAGER User turn DR DR i-1 , S i-1 A 1i A 2i MLP 1 MLP 2 (frames) System turn S AM i APPLICATION MANAGER First phase (MLP 1): Select the best request to be made to the AM . � Second phase (MLP 2): Generate the final system turn. �

:1*1$%,;&1$(7, Statistical Dialog Management & Z)$2,\9T&"(2@$0,206(,&1)(&0@,&%,-$54&)1&%$"*)5&-A-0,=-&

:1*1$%,;&1$(7, 4%6%+*%+#6/&-+6/05&I6$65"2"$%& Z)$2,\9T&%)26=,40-&&0)&$=/*,=,40&0@,&-A-0,=&/()=/0-&

89%/+$"& !" #$%&'()*%+'$, -" ./'01$,2+34'5,.67%187, F; !"#"$%&'("$)*G&?)61%6%+0$& 3> .%3%+7%+*34,(+34'5,83$35181$%, JH K*"(&I0)"/+$5& *> B8'%+'$,:1*'5$+%+'$, (> C'$%1D%E3F3&1,767%18, 1> A)4%+8'(34,#$%1&3*%+'$, G" =//4+*3%+'$,('83+$7, H" C'$*4)7+'$7,

:1*1$%,;&1$(7, To learn a Statistical Dialog Model Training Dialogs Dialog System Dialog Manager ??? Statistical Dialog Model

:1*1$%,;&1$(7, K*"(&I0)"/+$5&6$)&4+29/6%+0$& 1) Wizard of Oz technique � Pros: Acquisition under control conditions, not system fully developed. � Cons: Difficult to train the wizard(s) and to recruit people. DIHANA project: � We acquired a corpus of 900 dialogs. � We use a complete dialog system and the WOz simulates the role of the DM. � If we want to develop a complete statistical DM: � � To learn when a confirmation should be done � confidence measures.

:1*1$%,;&1$(7, K*"(&I0)"/+$5&6$)&4+29/6%+0$& EDECAN project: Two wizards of Oz: � Information Dialog Strategy System � ASR / NLU. History � DM. Speech Query Frame Dialog Appli c a t ion Und e r st anding Manag e m e n t Manag e r S i m ula t or S i m ula t or AM Answer System Frame TT S & An s w e r Graphi c al Answer Codified G e n e ra t or In t e r f a ce Answer 4.- Acquisition of a dialog corpus

:1*1$%,;&1$(7, K*"(&I0)"/+$5&6$)&4+29/6%+0$& 1) First wizard: UNDER S TANDING S IMULATOR Error & Und e r st anding Con f id e n ce M e a s ur e Edi t or S i m ula t or Correct Speech Simulated frame Frame 2) Second wizard: Only supervises the output generated by an initial dialog model. �

:1*1$%,;&1$(7, K*"(&I0)"/+$5&6$)&4+29/6%+0$& 2) Statistical user modeling

:1*1$%,;&1$(7, K*"(&I0)"/+$5&6$)&4+29/6%+0$& Formal description: � If the most probable user answer U i is selected at each time i , the � selection is made using the following maximization: All the possible user answers. We establish a partition in the space of sequences of states: � UR i : User register at time i (concepts and attributes). - Two different sequences of states are considered to be equivalent if they lead to the same UR i

:1*1$%,;&1$(7, K*"(&I0)"/+$5&6$)&4+29/6%+0$& Formal description: � The selection of the best U i is given by: � We propose using a MLP to make the assignation of a user turn: � the input layer receives the current situation of the dialog. � the output layer can be viewed as the a posteriori probability of � selecting the different user answers.

:1*1$%,;&1$(7, K*"(&I0)"/+$5&6$)&4+29/6%+0$& User Register representation: � The UR is a sequence of 15 fields: � Five concepts: Hour , Price , Train-Type , Trip-Time , and Services . � Ten attributes: Origin, Destination, Departure-Date, Arrival-Date, Departure- � Hour, Arrival-Hour, Class, Train-Type, Order-Number, and Services. We have assumed that the exact values of the attributes are not � significant to determine the next user answer: 0: The concept is not activated, or the value of the attribute is not � given. 1: The concept or attribute is activated. � The error simulator perform error generation and the addition of confidence � measures.

:1*1$%,;&1$(7, K*"(&I0)"/+$5&6$)&4+29/6%+0$& First, we evaluated the behavior of the original DM that was learned using the � training corpus (obtained by WOz). Then, we evaluated its evolution when the successful simulated dialogs were � incorporated to the training corpus. We defined four measures to evaluate the evolution of the dialog manager: � #unseen : the number of unseen situations. � #error : the number of answers provided by the DM that would cause the � failure of the dialog). %strategy : the percentage of answers provided by the DM that exactly � follow the strategy defined for the WOz to acquire the training corpus. %coherent : percentage of answers provided by the DM that are coherent � with the current state of the dialog.

:1*1$%,;&1$(7, K*"(&I0)"/+$5&6$)&4+29/6%+0$& #unseen

:1*1$%,;&1$(7, K*"(&I0)"/+$5&6$)&4+29/6%+0$& #error

:1*1$%,;&1$(7, K*"(&I0)"/+$5&6$)&4+29/6%+0$& %strategy and %coherent

:1*1$%,;&1$(7, K*"(&I0)"/+$5&6$)&4+29/6%+0$& 3) Automatic Dialog Generation: Interaction of a user simulator and a dialog manager simulator. � Initial model: Random selection of one of the possible answers defined for � the semantics of the task (user and system dialog acts). Learning: the probabilities of the answers selected by the dialog manager � during that dialog are incremented before beginning a new simulation. An error simulator module has been designed: � � Performs error generation. � Addition of confidence measures. Evaluation: A set of stop conditions is defined to automatically evaluate is � the dialog is successful or not.

:1*1$%,;&1$(7, K*"(&I0)"/+$5&6$)&4+29/6%+0$& Automatic Acquisition of a Dialog Corpus: � USER SIMULATOR DATABASE QUERY Databases MANAGER USER SEMANTICS User Frames NATURAL DIALOG MANAGER LANGUAGE UNDERSTANDING AND ASR SYSTEM SEMANTICS GENERATION ERROR SIMULATOR User Frames with System Errors and Frames Confidence Scores DIALOG HISTORY AND STOP CONDITIONS User Turn - System Turn

:1*1$%,;&1$(7, Example of a dialog: � S1: ( Opening ) Objective: Welcome to the sport service. How can I help you? Booking U1: ( Booking ) [0.9] Sport: tennis Sport : Tennis [0.9] Date : 03-15-2011 [0.1] Date: tomorrow Hour : 08.00-09.00 [0.9] S2: ( Confirmation-Date ) Do you want to play on the 15th March? U2: ( Negation ) [0.9] Date : 03-14-2011 [0.9] S3: ( Confirmation-Booking ) Sport : Tennis Date : 2011-03-14 Hour : 08.00-09.00 {One court available} Do you want to book tennis court number 2? U3: Yes. S4: ( Booking ) Sport : Tennis Date : 2011-03-14 Hour : 08.00-09.00 ( New-Query ) Tennis court number 2 has been booked. Anything else? U4: ( Negation ) [0.9] S5: ( Closing ) Thank you for using the sport service. Goodbye.

89%/+$"& !" #$%&'()*%+'$, -" ./'01$,2+34'5,.67%187, F; !"#"$%&'("$)*G&?)61%6%+0$& 3> .%3%+7%+*34,(+34'5,83$35181$%, ?> @71&,A'(14+$5, #H L20%+0$&!"#05$+%+0$& (> C'$%1D%E3F3&1,767%18, 1> A)4%+8'(34,#$%1&3*%+'$, G" =//4+*3%+'$,('83+$7, H" C'$*4)7+'$7,

:1*1$%,;&1$(7, L20%+0$&("#05$+%+0$& � .GG,0+"C,'E%-/4+"1& E& � GA-0,=&0@"0&"(,&"C*,&0)E& � ������������������������������ & � .%"/0&0@,&)/,("0$)4&"22)(%$45&&0)&0@,=;& & � W()#$%,&,=)0$)4"*&(,-/)4-,-E& � �������������������������������������� & � 3=/()#,&@6="4I2)=/60,(&$40,("20$)4;& & & &

:1*1$%,;&1$(7, L20%+0$&("#05$+%+0$&

:1*1$%,;&1$(7, L20%+0$&("#05$+%+0$& The emotion recognizer � Based solely in acoustic and dialog information: � firstly takes acoustic information into account to distinguish � between the emotions which are acoustically more different, secondly dialog information to disambiguate between those � that are more similar. Recognizing negative emotions that might discourage users from � employing the system again or even lead them to abort an ongoing dialog: anger , boredom and doubtfulness . Our emotion recognizer employs acoustic information to � distinguish anger from doubtfulness or boredom and dialog information to discriminate between doubtfulness and boredom .

:1*1$%,;&1$(7, L20%+0$&("#05$+%+0$&

:1*1$%,;&1$(7, L20%+0$&("#05$+%+0$&

:1*1$%,;&1$(7, L20%+0$&("#05$+%+0$&

:1*1$%,;&1$(7, L20%+0$&("#05$+%+0$&

:1*1$%,;&1$(7, L20%+0$&("#05$+%+0$&

89%/+$"& !" #$%&'()*%+'$, -" ./'01$,2+34'5,.67%187, F; !"#"$%&'("$)*G&?)61%6%+0$& 3> .%3%+7%+*34,(+34'5,83$35181$%, ?> @71&,A'(14+$5, *> B8'%+'$,:1*'5$+%+'$, )H M0$%"@%N6B6("&*7*%"2& 1> A)4%+8'(34,#$%1&3*%+'$, G" =//4+*3%+'$,('83+$7, H" C'$*4)7+'$7,