data science weekend 2017. segmento, На пути к идеальной диалоговой...

21
programmatic-платформа, группа компаний Сбербанк www.segmento. ru На пути к идеальной диалоговой системе Почему это так сложно и где мы сейчас programmatic-платформа, группа компаний Сбербанк www.segmento. ru

Upload: newprolab

Post on 11-Apr-2017

32 views

Category:

Data & Analytics


0 download

TRANSCRIPT

programmatic-, www.segmento.ru

programmatic-, www.segmento.ru

-:

( ):

(chit-chat):123

( ) + 2

:

:

5

, :

:

7

. :

-

. .

9

123

, ,

.

4

10

1. , u2. ()3. - , 4. - , (-1, 1)

, 11

: ,

Automatic SpeechRecognition (ASR)Spoken LanguageUnderstanding (SLU)Dialog State Tracker (DST)Dialog PolicyNatural LanguageGeneration (NLG)Text to Speech(TTS)

13

: rule-based

14

2

1 pos-tagger syntactic parser

3Coreference resolution

4 :

5

6 rule-base 7 DSTC4 .

15

LSTMSlot TaggingSigmoidLSTMLSTMLSTMLSTMLSTMLSTMLSTMNLUIntentsIntentsIntentsIntentsSlot TaggingSlot TaggingSlot TaggingNLUNLUNLUutter J-2 utter J-3 utter J-1 utter J System Actions at J +1LSTMLSTMLSTMLSTMLSTMLSTMLSTMLSTMLSTMLSTMLSTMLSTMSoftmaxSoftmaxSoftmaxSoftmaxEmbeddingEmbeddingEmbeddingEmbeddingSigmoidtag 1tag 2tag 3tag 4Slot taggingMulti-label intetsW 1W 2W 3W 4Shared weights

Natural Language Understanding block (NLU)

End-to-end joint learning of natural language understanding and dialogue manager

16

End-to-end :

17

WordEmbeddingCNNGRU2SimMatriesGRU1ScoreLMatchingAcclumulationPoolingConvolutionM , 12

Sequential Match Network

18

LSTMLSTM

?< S >< /S > reinforcement learning (Adversarial Learning)

, (seq2seq):

Adversarial Learning for Neural Dialogue Generation

19

,

(hierarchical encoder with binary classifier on the top):

Policy gradient training: J() = Eyp(y|x)(Q+({x, y})|)

softmaxsentenseattentionsentenseencoderwordencoderwordattention

,6Digital ctober [email protected] Vostryakov at linkedin