SDNet: Contextualized Attention-based Deep Network for Conversational Question Answering

Transcript

1. L 2 AA #  
    M )- A ( A EA : AA A ) - AD ( C A : A D 2
2. #1. Introduction
3. 
    3 -
4. . I - DI MI CAN ,II DIA D : H . I G G - D GH IA D C 4 HIA D ,DH GAD - D D 8 2A C 8 D D 0 D 2A G H I D .A C H G G G A . ( R HI I I GI Q X - 4, S ) 1DIG IA D 
   
5. 
   
6. lan 1 7 P P ztT o O lan P P 2- M ! , ((( I P T u ) + ((( I o O P M s e fr A 7 P oO z gT p h i - 7 0 8 QkuM cm 7 d ) .
7. Passage QA
8. Passage QA
9. Passage QA
10. Passage QA C ! A !
11. #2. Method
12. Me D P 8 + 0 ,. 1 2 ( ) à df l le lG Me ! "# f $# y G l A ( lQ "# 8 2 a eN 0f Me G tR oA rm 0 th G D l lf <Q> h <A> eu "* , $* , ", , $, , … , "#.* , $#.* , "# → $# "# → $# Reformulate!
14. I) B I B ) I( C C ) ) 2 I C B C B I C B -- I B E CC C I C B CC C I CC C I. CA C B I A B C C
15. BD ) 0 B 3 B 8 ( -B G8 3D D8 , 4 . 8 ( 2B 0D8 , 4 o E 8 E a 8 Dh Vvd puti M Tgo D CD E 8 B m ,3 r b W TgR o sPm v Tgo D CD E 8 B d Tg !o D CD E 8 B l no e Nn ( 1 B BERT& = ( )*+ , -) ∑/*+ 0 ℎ/ ) 2 , where 3 is # of layers and 2 is # of tokenized BPE tokens -) is weight of 4-th layer
16. 
    
17. - - - M / Q . / - - S R L B NT 2 / - - -
18. - - - ))R N C 2 - 2 - !" - - - - !! - - - - - - - N C - R N L - - -. # S - $ % & %'( ) à MV - - - - - - - ! ( - - -
19. 
    
20. - 2 - . . - !""# $, &, ' = !""#( ℎ+ , +-. / , ℎ + 0 +-. 1 , ℎ + 0 +-. 1 ) M 2 - C Q 2 - A , / A - - ; C
21. . B - B B . . . . . . . .2 ,
22. - M Q ( - )) CW - H - .2- N R H - Q 2 Q . .- 2 RN ( 2 - -
23. l - 2 3 1 - - l Q e - 2 Q 2 H ! ,=2 H " M l . e - 2 3 112 2Q 1 2 C 2 2h l NC 2 h HoW& ' &() * ∈ ℝ*×./ ,=2 h HoW & 0 &() 1 ∈ ℝ1×./ Q W 2 2 2 ∈ ℝ*×1Q i l 2 ∈ ℝ*×1 ℎ & 0,5 &() 1 ∈ ℝ1×.Q W d = = a ! & (5),' &() * Q i l c ! R h k 2 ,=2 2 2 2 2
24. - C R N 1 1 ( 1 . 1 ++ . ++S 1 1 1 . ) 1 S -1 2 . 1 - A M ++ !" !! 1
25. 
    
26. , . , . . . . . . Q MC A 2
27. 2 1 5 / .2 + M Ud / 2 S iNY Y aef :215 5 M cLh , /
28. #3. Experiments
29. - 3 C A E A Q . O C A 3. Out-of-domain
30. . 1 2 B c . 5 2 L T E T eR . - 32 . 2 E % - 32 F aa 2 627
31. ( r i d s i u m N p d u m o E D r d % 3 s u m r i 23 2 N n t u d u m ! = 0 ! ≥ 0 Se )72 3. Previous % round
32. #4. Conclusion
33. F D HDE DD D 4 C A EB D B D s . CC y F BC D CD B D B DD D O C DD D c e l c E n F BC D D C D xl CD D D BD m v t -A x E D DEB ui k & & c d P Q bl g SR s N r a . D x fh w B AB C D D DEB op T fh z i n N M EC & &
34. 724 / 1 D N 8 O 0 D D 98I 2 4 A 1 D I8 D8 I D 0DI D
     Q724 8 / I/ 8 N 89I - ( ) .) Q724 5 : 8 3C C D 8 D/ I/ 9 : C C : I 724 Q0 I I D 8 D C8 8 I / I/ D I 8D I C D8 8 I :O Q1 0 8 / I/ 8 N 89I - - ( Q1 0 C 8 / I/ I 8D D 9 : 8