An Auto-Encoder Matching Model for Learning Utterance-Level Semantic Dependency in Dialogue Generation

弊研究室で行なったEMNLP2018読み会の発表資料です。

Transcript

1. An Auto-Encoder Matching Model for Learning Utterance-Level Semantic Dependency in Dialogue Generation 
   1  1
2. 
   • ',"! %,(* +  • $     Seq2Seq
     • ',& )#  2
3. 
    
    !" # = − log ) * + +; #) !. / = 1 2 2 ℎ − 4 !5 6 = − log ) * 7 7; 6) !8 #, /, 6 = − log )(7|+; #, /, 6) 3
4. Auto-Encoder Encoder 
    !" # = − log ) * + +; #) !. / = 1 2 2 ℎ − 4 !5 6 = − log ) * 7 7; 6) !8 #, /, 6 = − log )(7|+; #, /, 6) 4
5. Mapping Module 
    !" # = − log ) * + +; #) !. / = 1 2 2 ℎ − 4 !5 6 = − log ) * 7 7; 6) !8 #, /, 6 = − log )(7|+; #, /, 6) 5
6. Auto-Encoder Decoder 
    !" # = − log ) * + +; #) !. / = 1 2 2 ℎ − 4 !5 6 = − log ) * 7 7; 6) !8 #, /, 6 = − log )(7|+; #, /, 6) 6
7. End-to-End train loss 
    !" # = − log ) * + +; #) !. / = 1 2 2 ℎ − 4 !5 6 = − log ) * 7 7; 6) !8 #, /, 6 = − log )(7|+; #, /, 6) 7
8. 
    •  Daily Dialogue Corpus 
   36.3k pairs  
   11.1k pairs 
   11.1k pairs •   BLEU-(1, 2, 3, 4), Distinct-(1, 2, 3), Human evaluation •   Seq2Seq, Seq2Seq + Attention 8
9. 
   •   •
      9
10. 
    10 • AEM 
      •     10 
11. 
     • *.% # Seq2Seq& • Seq2Seq+ • ( $,! • ( Fluency, Coherence   ' " • Seq2Seq  -)
    11