Share
Embed
Transcript
39 slides
-
스캐터랩(ScatterLab) ੌ࢚ച ੋҕמ 핑퐁팀 ML 세미나: Dialog System 김준성 Not
only Retrieval But also Generation Machine Learning Engineer -
!3 "An Adversarial Approach to High-Quality, Sentiment-Controlled Neural Dialogue Generation.”
Kong, Xiang, et al. arXiv preprint arXiv:1901.07129 (2019). য়ט ࣗѐೡ ֤ޙ #Dialog-Generation #DialogSystem #CVAE #CGAN -
!4 ಽ۰Ҋ ೞח ޙઁ ࢎਊ ച History ( N-Turn) ী
ೠ ೠ ߸ਸ Generation ೞ! ױ! хਸ ઑ೧ࢲ ߸ਸ Generation ೡ ࣻ ਸө? : ݍয? -> (ӛ) ! ݍয / (ࠗ) ইפ ߹۽য -
!5 ಽ۰Ҋ ೞח ޙઁ ࢎਊ ച History ( N-Turn) ী
ೠ ೠ ߸ਸ Generation ೞ! ױ! хਸ ઑ೧ࢲ ߸ਸ Generation ೡ ࣻ ਸө? : ݍয? -> (ӛ) ! ݍয / (ࠗ) ইפ ߹۽য Q. যڌѱ ࢎۈэ Quality ߸ਸ ٜ݅য յ Ԅؘ? -
!6 ಽ۰Ҋ ೞח ޙઁ ࢎਊ ച History ( N-Turn) ী
ೠ ೠ ߸ਸ Generation ೞ! ױ! хਸ ઑ೧ࢲ ߸ਸ Generation ೡ ࣻ ਸө? : ݍয? -> (ӛ) ! ݍয / (ࠗ) ইפ ߹۽য Q. যڌѱ ࢎۈэ Quality ߸ਸ ٜ݅য յ Ԅؘ? Q. যڌѱ ౠ хী ೠ ߸ਸ ٜ݅য յ Ԅؘ? -
DEEP-DIAL Workshop(at AAAI) The Second AAAI Workshop on Reasoning and
Learning for Human-Machine Dialogues -
!8 Motivation Sentimentח Dialog Systemী যࢲ ࢎۈٜ Engaigingҗ Interest ࢚थਸ
݅ٞ ী ೠ ৈ۞ दبٜ ݅, ࠗ࠙ rule-based ژח template ӝ߈ ߑधਸ ࢎਊೣ → ח જ ߸ਸ ղӝ ਤ೧ࢲח ݆ మ݁җ ܙٜ ਃ۽ ೞח ޙઁ End-to-End ӝ߈ Generation ߑߨٜ Dialog System োҳীࢲ ਃೠ ઁ۽ ځয়ܰҊ → Reason: Rule-Base ࠁ ഻न ਬোೞҊ ࢎਊೞӝ ਊೠ ߑߨী োҳ ೯ -
!9 Previous Research ୡӝ োҳীࢲח SEQ2SEQ ӝ߈ ݽ؛ٜ ݆ ࢎਊ
غਵա, റ ؊ જ ௬ܻ౭৬ بೠ ߸ਸ յ ࣻ ب۾ ݽ؛ ࢿמ ೱ࢚ਸ ਤೠ ৈ۞ োҳٜ ܖয . ؘఠܳ ӝ߈ਵ۽ Dialog End-to-End Modelীࢲ Sentiment ࠁܳ Replyী ߈दః۰ח ݻݻ োҳٜ द ઓ೮חؘ, ৈ۞ ޙઁٜ ز߈೧ࢲ ߊࢤ ೮ 1.Hasegawa(2013): п emotion݃ ܲ ݽ؛ਸ णदఇ → ষդ training-cost, sparsity 2. Shen et al (2017): latent variableী emotion infoܳ ֍݅, response хਸ যڌѱ ஶ܀ೡ x 3.Zhou and Wang (2018): twitterী ח emojiо ೧ contextܳ ߈ೡ Ѫۄ оࢸਸ ࣁҊ ݽ؛ ࢸ҅, Cond Variational AutoEncoderਸ ਊ೧ࢲ Sentimentܳ Control ೡ ࣻ ח dialog-system ѐߊ -
!10 Previous Research 1. Hasegawa(2013): п emotion݃ ܲ ݽ؛ਸ णदఇ
→ ষդ training-cost, sparsity 2. Shen et al (2017): latent variableী emotion infoܳ ֍݅, response хਸ যڌѱ ஶ܀ೡ x 3. Zhou and Wang (2018): twitterী ח emojiо ೧ contextܳ ߈ೡ Ѫۄ оࢸਸ ࣁҊ ݽ؛ ࢸ҅, Conditional Variational AutoEncoderਸ ਊ೧ࢲ Sentimentܳ Control ೡ ࣻ ח dialog-system ѐߊ -
!11 Proposal ֤ޙীࢲח Conditional Generative Adversarial Network (CGAN)җ
Conditional Variational Auto Encoder (CVAE) ӝ߈ਵ۽ хਸ ઑೡ ࣻ ח ച ࢤࢿ ݽ؛ਸ ݅ٞ Fluency৬ Controlable فܻ݃ షՙܳ ӝ ਤ೧ࢲ فѐ ஹನք۽ ա־যઉ : Generator, Discriminator Generatorח History৬ Emotion Labelਸ ਊ೧ࢲ ߸ਸ Generation ೞҊ Discriminatorח History, Emotion Label, Responseܳ ࠁҊ ѱ Ground Truthੋ ݏ୶ѱ ण(դ) Generatorо Discriminatorܳ ҅ࣘ ࣘب۾ णೞݶࢲ োझۣѱ Response ௬ܻ౭ܳ ֫ -
VAE: Variational Auto Encoders Input X Reconstructed X Input Xܳ
ౠ Vector(Z: latent variable)۽ Abstraction दఃҊ ܳ द Reconstruct Encodeীࢲ աৡ Representationী ೠ ࠁ Ө ೧ -
Conditional GAN | VAE GANҗ VAE۽ ࢤࢿػ Ѿҗܳ ܻо Control
ೡ ࣻ হ GANҗ VAE ઁੌ ޙઁ : ࢎۈ ҷ Input ->Attribute(Ѩ࢝) ॆӖۄझܳ ՛ ࢎۈਸ ۱೧ -
Conditional GAN | VAE Ӓ۞ݶ GAN, VAEܳ णदఆ ٸ ౠب
э ֍ਵݶ غ ঋਸө? Conditional GAN, VAE Motivation -
Conditional GAN | VAE https://arxiv.org/abs/1702.01983 Antipov, Grigory, Moez Baccouche, and
Jean-Luc Dugelay. "Face aging with conditional generative adversarial networks." 2017 IEEE International Conference on Image Processing (ICIP). IEEE, 2017. "Face aging with conditional generative adversarial networks." -
Conditional GAN | VAE https://arxiv.org/abs/1702.01983 Antipov, Grigory, Moez Baccouche, and
Jean-Luc Dugelay. "Face aging with conditional generative adversarial networks." 2017 IEEE International Conference on Image Processing (ICIP). IEEE, 2017. -
Base Model Dialog SEQ2SEQ with Attention 1-layer bi-directional GRU (128-hidden)
Dialog History (Previous Utterances) -
Conditional Variational Auto Encoder 1. Utterance History Encoding -> Context
Vec C 2. Embedding Sentiment Label -> Sentiment Vec S 3. Concat Context + Sentiment Vector = Z 4. Decoding Reply Sequence with Z -
ળࢿ: ղо যઁ ޤݡח ঌই? : উҾӘ೧ ળࢿ: ইפ ೠߣ
ݏࠊ! Wh Dialog History Context Vec C + S = Z Latent Variable Conditional Variational Auto Encoder Encoder Wr : ঌѷয ޤ ݡחؘ? Generator -
ળࢿ: ղо যઁ ޤݡח ঌই? : উҾӘ೧ ળࢿ: ইפ ೠߣ
ݏࠊ! Wh Dialog History Context Vec C + S = Z Latent Variable Conditional Variational Auto Encoder Encoder Wr : ইפ উҾӘೞҊ! Generator -
Conditional Variational Auto Encoder https://hugrypiggykim.com/2018/09/07/variational-autoencoder%EC%99%80-elboevidence-lower-bound/ ੌ߈ੋ VAE Objective Function ҳઑ
| Minimize Dialog-CVAE Objective Function ҳઑ | Maximize -
Conditional Variational Auto Encoder https://hugrypiggykim.com/2018/09/07/variational-autoencoder%EC%99%80-elboevidence-lower-bound/ ߸ Input -> Latent Variable(Z),
х чਸ ݏۄ! Latent Variable(Z) -> ߸, х чਸ ݏۄ! فѐ Objectܳ زदী णೞח ߑध -
Conditional Variational Auto Encoder https://hugrypiggykim.com/2018/09/07/variational-autoencoder%EC%99%80-elboevidence-lower-bound/ : ইפ উҾӘೞҊ! Wr
х : Context Vector(Z) : ળࢿ: ղо যઁ ޤݡח ঌই? : উҾӘ೧ ળࢿ: ইפ ೠߣ ݏࠊ! х : Context Vector(Z) : ળࢿ: ղо যઁ ޤݡח ঌই? : উҾӘ೧ ળࢿ: ইפ ೠߣ ݏࠊ! : ইפ উҾӘೞҊ! Wr -
Conditional Generative Adversarial Network 1. Utterance History Encoding -> Context
Vec C 2. Embedding Sentiment Label -> Sentiment Vec S 3. Concat Context + Sentiment Vector = Z 4. Decoding Reply Sequence with Z 5. Discriminator(Decoded_output, Context, Sentiment) -
ળࢿ: ղо যઁ ޤݡח ঌই? : উҾӘ೧ ળࢿ: ইפ ೠߣ
ݏࠊ! Wh Dialog History Context Vec C + S = Z Latent Variable Encoder Wr : ঌѷয ޤ ݡחؘ? Generator Conditional Generative Adversarial Network -
ળࢿ: ղо যઁ ޤݡח ঌই? : উҾӘ೧ ળࢿ: ইפ ೠߣ
ݏࠊ! Wh Dialog History Context Vec C + S = Z Latent Variable Encoder Wr : ঌѷয ޤ ݡחؘ? Generator Discriminator Ѥ о ೠ ݈ੌө ইפݶ Generatorо ࢤࢿೠ ݈ੌө? Conditional Generative Adversarial Network -
ળࢿ: ղо যઁ ޤݡח ঌই? : উҾӘ೧ ળࢿ: ইפ ೠߣ
ݏࠊ! Wh Dialog History Context Vec C + S = Z Latent Variable Encoder Wr : ঌѷয ޤ ݡחؘ? Generator Discriminator Ѥ о ೠ ݈ੌө ইפݶ Generatorо ࢤࢿೠ ݈ੌө? Discriminatorܳ ࣘӝ ਤ೧ࢲ ؊ ۢ ঠӝ ೡకঠ Conditional Generative Adversarial Network -
ࢿמ: ಣо ױࣽೠ SEQ2SEQ ਵ۽ח ݈ ࢿמ ߹۽ જ ঋ
(ACC 55%, PPL 150+) ਵ۽ CVAEܳ ా೧ࢲ णػ Generation જ ࢿמਸ ࠁח Ѫ э (ACC 75%, PPL 81) CGAN݅ਸ ਊ೧ࢲח ѱ જਸ ࢿמਸ ࠁ ޅೣ -> ࢎۈਸ ղղח Ѫ ݅ਵ۽ח ൨ٞ (ACC 64%, PPL 120) -
ࢿמ: ಣо CVAEо ೧֬ ߏী CGAN ઑܐ פ SOTA Generation
ػ Replyী ೠ ࠁ Ө ೧ : CVAE ࠁ ࢎۈۢ ߸ೡ ࣻ ח ղղӝ : CGAN -
୨ಣ Context-Aware Pretrained BERT۽ ച ࢤࢿ పझܳ Transfer-Learning ೞݶࢲ
ࢤࢿػ ߸ੋ ইפݶ पઁ ചੋ ҳ࠙ೞח Discriminatorܳ ୶о೧ࢲ োझۣѱ ച ௬ܻ౭ܳ ֫ৈ рݶ ࢎۈ э ചܳ ࢤࢿೞח ࠈਸ ٜ݅ ࣻ ঋਸө?!?! → Conditionalਸ ୶оೞݶ Blackbox rate ੌ ࣻ ޥо ೧ࠁҊ र!! ࢎۈۢ ҳয۽ ೡ ࣻ ਸө??