SGM: Sequence Generation Model for Multi-Label Classification

弊研究室で行なったCOLING2018読み会の発表資料です。

Transcript

1. SGM: Sequence Generation Model for Multi-Label Classification 2018/10/23
     1
2. 1.
    2.  3.  4.   5.  2
3. 1.
    2.  3.  4.   5.  3
4. #" n Multi Label Classification(MLC) 
   (,$. >2 !'& -E n ?6: 
   MLC-E>2Single Label Classification)+ Binary Relevance
   pairwise ranking loss; $ ! 6: n 
   %0:  !C<D
   3B) ! *574
   =/  n A3 
   seq2seq; sequence generation! 18
   %@-E?9 4
5. 1.
    2.  3.  4.   5.  5
6.  
    6
7.  
    7
8. Encoder n Bi-LSTM  n $
   !  #  n $
   ! # "#  !" = LSTM !"() , +" !" = LSTM !",) , +" !" = !" ; !" 8
9.  
    9
10. Attention n *  &2#,.+
    * '% n Attention
    *( 3 4" /1  n !" , $" , %"
    &'-$( decoder!40) 10
11. Attention n 
    !   n  
    Decoder  11
12.  
     12
13. Decoder n LSTM  n %#" n !"#$
    % − 1 
    ! n ( !"#$ 
    % − 1 &!  global embedding($) 13
14. Decoder n $(& !%  )' n !" , !$ , %$ n &'
     !!   " #  &' ( = * −∞ ( .
     ) 0 12ℎ456.74 14
15. Global Embedding n  
    #!* % 
    ) (!* n #!*-+ 
    ".!* '&/ (exposure bias) n Global embedding $, n !" , !$ ∈ ℝ'×'  15
16. 1.
     2.  3.  4.   5.  16
17. "- n  l Reuters Corpus Volume I (RCV1-V2) l '800,000 (  l Arxiv Academic Paper Dataset (AAPD) l 55,840 )$ !* l 
    &,#.+  % 17
18.  n  l Hamming-loss l! ", $ " = & '()*+( ∑ -./ '()*+(0& 1("- ≠ $ "- ) l Micro-F1 n  
     l Binary Relevance(BR) l Classifier Chains(CC) l Label Powerset(LP) l CNN l CNN-RNN 18
19. 
    19
20. 1.
     2.  3.  4.   5.  20
21.   n Global Embedding  
    ! "#$%  
    &  n  21
22.  n   
     sorting  
    Ablation Experiment  n  22
23. !  23 n  "(  )  
      
24. 1.
     2.  3.  4.   5.  24
25. 9; n Multi-label classification"68&3/0
     =5> ( n 1 decoder2<4 sequence generation  4
    %@7,)# /' n
    *   ! +. 
    $-:? 25