Modeling Multi-turn Conversation with Deep Utterance Aggregation

弊研究室で行なったCOLING2018読み会の発表資料です。

Transcript

1. Modeling Multi-turn Conversation with Deep Utterance Aggregation M1 Koji Tanaka 1
2.  –,+! ' 2 %0#)-(0 – (-:47".10*+ –+/2
   %0 $ 38?5<?2 &;9>,  –AE-commerce Dialogue Corpus 2==@6B 2
3. 
   
    3
4. Matching Attention Flow !" = $%& !"'( , *" , +" ,- " = ./ tanh 4 5 *- + 47 5 *" + 89 :; " = exp ,; " / @ -A( B exp(,- ") +" = @ ;A( B :; "*;
    4
5. 
    –Ubuntu Dialogue Corpus (Ubuntu) –Douban Conversation Corpus (Douban) –E-commerce Dialogue Corpus (ECD) TRAIN VALID TEST Ubuntu 1M 500K 500K Douban 1M 50K 10K ECD 1M 10K 10K [context-response pairs] 5
6.   –
    – Rn@k –  – Multi-turn (Zhou et al, 2016) – Sequential Matching Network (Wu et al, 2017) 6
7. 
       7
8. EDC    
   8 
9.  –1   – 
    U: How about the packaging of skin care products. By the way, which delivery company will be responsible for shipping and how long can I receive the goods ? U:How about nuts ? S: Nuts is good. U: OK then, how about zongzi ? 9
10.  –!$*"'* &%)  –
      –,E-commerce Dialogue Corpus ((+#- –EDC  10