snlp2018slide

Personalizing Dialogue Agents: I have a dog, do you have pets too? (ACL2018)

Transcript

1. Personalizing Dialogue Agents: I have a dog, do you have pets too? [Zhang+, ACL’18] ) 2( NLP , 2018-08-03
    
    
2. .:&=- #)09'*( '*( l .:&=@> #)09,  !A  /;3Persona+ )"09 <27 %$ ?4 – 56, “I don’t know.” G$9?.C  [Li+, NAACL’16] 2 B856 1 [Li+, ACL’16] l 4%>7$2! " (Twitter)/
    4%1 embedding*49?(+0@ )- – FA!8&'43:4%D – 4%#,
   9?  <= &EB;
3. 
    l *&#"-!>^U8Persona ? (,'+M 9=  .D%$IA (PERSONA-CHAT Dataset) – ZGD1(,'+6QRK – D1Persona@ VW B7 .D%$O 3_J[ l \CQX;]
   QX5A)%+> PersonaLSQXK:P<YH – /4YH2TPersonaEQX 0NF 3
4. Proposal: PERSONA-CHAT Dataset l "# %
   745   Persona2+ (Original Persona) – $. 1,1550 l '5" &non-stop word * (3-8(Revised Persona) – ,)1
   /6 ! 4Persona2+ 4
5. Proposal: PERSONA-CHAT Dataset l )Persona $( 
   
    %' – “*  #  ” &" l ! 10,981dialogs (= 164,356 utterances) 5
6.   Persona  l 7;8@WU (+ ,1+$0H) + MRTO – LH3<IG !    Q.)0 BN F:6@" 8@WU?  … l 4;MRTO8 &%# – (&*9>CMR + /2'-=V 19MR5A20MRTO!!&%#"KP – &%#"DMR"1=V >CMR"=VJ5 Hits@1 SE 6
7.    l Ranking Models: – IR Baseline [Sordoni+, ’15] – StarSpace [Wu+, ’17] – StarSpace + Profile Memory Network [Sukhbaatar+, ‘15] – StarSpace + Key-Value Profile Memory Network [Miller+, ’16] l Generative Models: – Seq2Seq [Sutskever, ’14] – Seq2Seq + Profile Memory [Sukhbaatar+, ‘15] 7 Seq2Seq 
   perplexity   q c  
8. Ranking Models (1/3)  +( #%'$ q, c    !"&*  ) l IR Baseline [Sordoni+, ’15] – JE tf-idf=@ Bag-of-Words 0<QNAGKC 69 l StarSpace [Wu+, ’17] – Word embedding>;0<QNAGKC 69 – margin ranking loss 48I3embedding 5M 8
   (+'!* O/.0<QN)#&*q “concatenate”  ?AGKC)#&*c- L  $%",HP27… “add”? 1D q ⌃imax (0 , µ sim ( q, c ) + sim ( q, ci )) : :F/BFAG c, ci : 0<QN
9. Ranking Models (2/3) l StarSpace + Profile Memory Network: –   q Persona"p i $  # % (= q+) 9 q : 
   " : !   pi q + = q + ⌃ sipi, si = Softmax ( sim ( q, pi)) arXiv& fig.1 
10. Ranking Models (3/3) l StarSpace + Key-Value Profile Memory Network [Miller+, ’16]: – Profile Memory NetworkF> q+"8 GD(&/CH 29
    @B*&. 529+$)-q++"<? – #'.%,.;" u j <? , 7=A 0! +$)- u j 3 @B r j "8 – CH 29IE u j "1   3  @B r j H6   r j 4"q+ : 10 uj : GD(&/ 39IE
    @B rj q ++ = q + + ⌃ s 0 jrj, s 0 j = Softmax ( sim ( q + , uj))
11. Generative Models 11 l Seq2Seq [Sutskever, ’14] – $)/+%(Seq2Seq3-5 ESW<H F Y ,(.APerplexityGD SW C\ – 160"5P E;:>BPNU_  LSTM9^#7&/ (!?) l Seq2Seq + Profile Memory Network – 8Ranking Model@Profile Memory Network = Profile2%*? !,7'47 – 160"5#7&/V@X]embedding OI !,7'47LMQ V@R T ZJK[
12. 1-1: Ranking vs Generative l Generative Model % $' – Generative Model"% # & 
     !   12
      Original Persona / Revised Persona  Persona ( Revised
    
    )
13.   1-2 : Persona 
      (Original) l Original Persona  hits@1 – Seq2Seq     13
14.   1-3 : Persona 
      (Revised) l Original Persona%4hits@1/5 – Seq2Seq 3 $()+- l Revised Persona* '&No Persona 1 – !# "0.2,6  14
15.    l ),$'Persona?D  ") &+!(%*4-:0 – Ranking ModelPL%#.AG 5N Generative ModelAGJEJ;17 – Personarevise 2(Original Persona)> l -:@M3") &+! – Fluency, Engagingness, Consistency 3O6 5FIH= – -: (%*?D Persona 2KC 5N9<B(%* Persona Q-:/8  H= 15
16. 2-1:  Persona 16 l 
    Persona Detection rate   –  
      ( )
    
17. 2-2: 3 17 l !
    + '%Persona Detection rate ( – &1!
    3 "-$,  l 0#*/)-  2(.   ( )
    
18.  
     (Ranking) (P1: Human, P2: KV Profile Memory) 18
19.  
     (Generative) (P1: Human, P2: Profile Memory) 19
20.  l #%$ l +*97 –  +@:% Persona, 6  &*!03. (PERSONA-CHAT Dataset) – >/ 8;(?
    8;$.+ Persona 59 8;4' 7)<2 l "1'( –   6Persona  +#"8;(? -4 1= 20  8 )4/Persona &   .3 5-2 ,3 ! 60
21. 
    l 2.3@YF= g_<f>8aE W?CU"Appendix]!$" b _V_V6Q l G@SNS!dJTM04/'3& A<HS^*+-1X  – 04/'3&OA   .)(5 $ 7H., !R 04/'3 ce# N%;`  – !@I7H04/'3BK $ :H9_?L>DP
    [ S^Z\ # 21
22.  
      
23. 2-3: 
         23 l !'
      TwitterOpensubtitles$ ., Seq2Seq )/ PERSONA-CHAT .,  0+" – TwitterOpensubtitles ”-( !'”  
       &# %* 
24.  3-1:  Persona 
     l (Ranking Model+. l 16 *'#Persona (Their Persona)  & $!0/4+., – 
    "%'# -23 7 - * )5)5 24
25.  3-2: Original Persona
     l Original Persona 30"!-4'*.% – 12 Revised Persona& 30 -4 
    +)5*$(   –  Revised Persona& "! , No Persona #/ 25