Share
Embed
Transcript
32 slides
-
ݾର ݾର! 1. ޙઁ ࣗѐ : যڃ ޙઁܳ ಽ۰Ҋ ೡөਃ?
2. ޙઁ ࢸ : ؘఠࣇ / పझ / Single vs Multi-Attribute 3. ಣо ߑߨ : Style Transfer Degree, Content Preservation, Naturalness 4. ݽ؛ : Disentanglement vs Entanglement -
Style Transfer ঌইࠁӝ ޙઁ ޙীࢲ झఋੌۄח Ѫ ޖੌөਃ? द)
• ޙ ӛ, ࠗ • ޙয, ҳয, ࢿ҃(?) • যൃ ࢶఖ (য۰ ਊয, ए ਊয) • ݈ై (যܲझ۞ ݈ై, গੋՙܻ ݈ై) -
Style Transfer ঌইࠁӝ ޙઁ Text style transfer is the
task of transferring the style of text having certain stylistic attributes, while preserving non-stylistic or content information - Papers With Code ੑ۱ ޙਸ ౠ झఋੌਸ ನೣೠ ޙਵ۽ ߸҃ೞח ޙઁੑפ. ױ, ߸҃दী ޙ ஶబஎ ژח ޙҗ ҙ۲ হח ࠁٜਸ ࠁઓ೧ঠ פ • ) ӛੋ ܻ࠭ -> ࠗੋ ܻ࠭ Text Style Transfer ۆ ޖੌөਃ? -
Style Transfer ঌইࠁӝ ޙઁ • Low Resource : Supervised
Learning ਵ۽ ಽӝীח Parallel ؘఠо ࠗೣ • (ੑ۱-߸ചػ ޙ) हਸ ٜ݅ӝ য۵Ҋ, ݅ٚҊ ೞ؊ۄب ࢎۈ݃ ҙ - Evaluation: ݺഛೠ ਸ ೞӝ য۵Ҋ, ࢎۈ ಣо৬ োҙࢿ ֫ ܳ ࢸೞӝ য۰. - Content Preservation: ޙ झఋੌ ߄Շݶࢲ ӝઓ ஶబஎо ߸ഋغয ߡܻח ޙઁо - ݽഐࢿ : झఋੌۄח Ѫਸ ݺഛೞѱ ೞӝ য۵ӝ ٸޙী, ಽযঠೡ ޙઁо ݽഐೣ పझীࢲ ೧Ѿ೧ঠ ೡ ޙઁ -
Style Transfer ঌইࠁӝ YELP Review Dataset - Positive(266,041), Negative (177,218)
https://www.yelp.com/dataset Amazone Review Dataset - Positive(64,251,073), Negative (10,944,310) Single Attribute п ܻ࠭ী ೠ ӛ/ࠗ ۨ࠶݂ غয ח ؘఠࣇ IMDb - Positive(178,869), Negative (187,597) -
Style Transfer ঌইࠁӝ FYELP (Multi-Attribute) YELP ؘఠࣇী Gender, Category Attribute
ܳ ୶оೠ ؘఠࣇ (MULTIPLE-ATTRIBUTE TEXT REWRITING, ICLR 2019) -
Style Transfer ঌইࠁӝ Empathetic Dialog Dataset (Multi-Attribute) ೧ ച х
ۨ࠶݂ غযח ؘఠࣇ. х -> ܲ хਵ۽ ߸ചदఃח ਊب۽ ࢎਊ оמೣ (Towards Empathetic Open-domain Conversation Models: a New Benchmark and Dataset, ACL 2018) (Zero-Shot Fine-Grained Style Transfer, arxiv 2019) -
Style Transfer ঌইࠁӝ Transfer ػ ޙ ݾ۽ ೞח झఋੌী ݃ա
Ӕೠо? ఫझ झఋੌਸ ߄Բח ӝࠄੋ ݾо ׳ࢿ غח ഛੋೞח Style Transfer Intensity ಣо ߑߨ - Style Classification (: Sentiment Classification) - Style Representation Similarity : Ӓ ച ѐԵ੨! ( ޙ pos ੋо?) <- (input) Ӓ ച ֢੨য, (target style) pos -
Style Transfer ঌইࠁӝ ӝઓ ޙীࢲ Ҋ ؍ ஶబஎܳ ࠙ ࠁઓ೮חо?
Style Transferܳ ೞݶࢲ ޙ ஶబஎо ࢚ࣚغѢա ߸ഋ غחܳ ഛੋೞח Content Preservation (=Relevance) ಣо ߑߨ - BLEU, METEOR, Overlap - Representation Similarity - Sentence Matching Model : Ӓ ച ѐԵ੨! ( ޙ inputҗ زੌೠ ஶబஎ?) <- (input) Ӓ ച ֢੨য, -
Style Transfer ঌইࠁӝ Naturalness (=Fluency) ࢤࢿػ ޙ োझ۞ ޙੋо?
যࣽ ੜޅغযب content preservation, style transfer intensity ח ֫ਸ ࣻ ಣо ߑߨ - Language Model Perplexity - Adversarial Evaluation (ѱ ࢎۈ ॵ ޙੌө ࠈ ॵ ޙੌө?) : Ե੨! Ӓ ച, ଢ଼ ( ޙ ઁ۽ ࢤࢿ ⪟פ?) <- (input) Ӓ ച ֢੨য, -
Style Transfer ঌইࠁӝ Tradeoff Evaluating Style Transfer for Text, 2019
NAACL Style Transfer Intensity ৬ Content Preservation, Naturalness ח Tradeoffо ࢤӡ ࣻ ߆ী হ -
Disentanglement о ަө? VAE Latent Representation Content Style Disentangled VAE
Latent Variable ܻח Style ਸ Control ೞҊ रؘ, Latent Representation Content ৬ Ԙৈ… ࢚క۽ ޥо Latent Representationܳ ߸ഋೠݶ contents ب ߸ഋؼ ށۄ! ഒઙ?! Content ৬ Style р Үਸ ୭ࣗചೞҊ, Style ݅ ઑೞ! -
Disentangled Representation Learning for Non-Parallel Text Style Transfer, ACL 2019
Disentangled Representation Learning for Non-Parallel Text Style Transfer, ACL 2019 - VAE ীࢲ ݅ٚ Latent Variable ਸ Style җ Content Latent Repre ۽ ܻ࠙ೞҊ ೞח Method - Style Repre ਸ ܻ࠙ೞӝ ਤ೧ࢲ Style-Classification, Content Repreܳ ਤ೧ࢲ CBOW objective - п Repreܳ ৮ ܻ࠙ೞӝ ਤ೧ࢲ п Term ী Adversarial Loss ܳ - ۧѱ ܻ࠙ػ Styleҗ Content Repre concat ೧ࢲ Decoding. Style Repre ݅ ߄Լ ֍ਵݶ Style Transfer غח ਗې -
Disentangled Representation Learning for Non-Parallel Text Style Transfer, ACL 2019
Disentangled Representation Learning for Non-Parallel Text Style Transfer, ACL 2019 -
Disentanglement о ަө? VAE Latent Representation Content Style Disentangled VAE
Latent Variable Ӕؘ Ѣ ೧ࠁפӬ ل ܻ࠙ೞӝ غѱ য۵…. ইפ ӒܻҊ গୡী ف ҕр ৮߷ೞѱ ܻ࠙ؼ ࣻ חо?! -
MULTIPLE-ATTRIBUTE TEXT REWRITING MULTIPLE-ATTRIBUTE TEXT REWRITING, ICLR 2019 - Style
Latent Representation ਸ ܻ࠙ೞח ߑߨ (Disentanglement Method) ਬബೞ ޅೞח Ѫ ਸ पਸ ా೧ ૐݺೠ ֤ޙ - Denoising auto-encoding ਸ ࢎਊ೧ࢲ Style Tokenਸ Masking ೞҊ ܳ Style Condition যਸ ٸ ೧ షਸ ݏ୶ח ޙઁ. - Style Token Classification ਃ۽ೣ - Multiple Attribute ޙઁܳ ಽ х -> ܲ х, ৈࢿ -> թࢿ, -> ֢ੋ - ֤ޙী Figure о হ…. -
Revision in Continuous Space: Unsupervised Text Style Transfer without Adversarial
Learning Revision in Continuous Space: Fine-Grained Control of Text Style Transfer, AAAI 2019 - ICLR 2019 Ѿҗܳ ߄ఔਵ۽ ѐߊػ VAE ߑध Disentanglement Method - ݽٚ ޙ Continuous Space ী Ҋ оೞҊ झఋੌ ߸҃ػ ޙਸ ইࠁ! - ߸ചػ Style Latent Variable ਸ ӝ ਤ೧ Predictor Error Surface ীࢲ Gradient Discent ܳ ࢎਊ. ী۞о ୭ࣗച غח ನੋо ܻо Ҋ ೞח ޙۄҊ оೞҊ ޙઁܳ ಽӝ - Predictor = VAE + Attribute Predictor + Content Preservation ೧Ѿਸ ਤ೧ CBOW بੑ -
Revision in Continuous Space: Unsupervised Text Style Transfer without Adversarial
Learning Revision in Continuous Space: Fine-Grained Control of Text Style Transfer, AAAI 2019 -
MULTIPLE-ATTRIBUTE TEXT REWRITING - AAAI 2019 ֤ޙҗ Entangled Latent Repre
ܳ Gradient ߑधਵ۽ ࣻೠח Ӕ ߑध ਬࢎ - णदীח VAE reconstruction loss৬ attribute classifier ܳ णೞب۾ פ. - encoder ۽ ࢤࢿػ latent variable ਸ ҅ࣘ সؘ ೞݶࢲ attribute classifier ৬ loss о ੌ threshold ೞо ؼ ٸ ө ߈ࠂೞݴ latent variable ਸ ࣻפ. Controllable Unsupervised Text Attribute Transfer via Editing Entangled Latent Representation, NeurIPS 2019 -
MULTIPLE-ATTRIBUTE TEXT REWRITING Controllable Unsupervised Text Attribute Transfer via
Editing Entangled Latent Representation, NeurIPS 2019 -
Style Transformer: Unpaired Text Style Transfer without Disentangled Latent Representation,
ACL 2019 Style Transformer: Unpaired Text Style Transfer without Disentangled Latent Representation, ACL 2019 - Style Latent Representation ਸ ܻ࠙ೞ ঋח ߑध (Entanglement ߑधਸ ࢎਊೣ) - Transformer ҳઑ Encoder, Decoder ܳ ࢎਊೣ - Self-Transfer(1), Back-Translation(2), Style Discriminator(3) ܳ زदী णೞח ҳઑ -
Style Transformer: Unpaired Text Style Transfer without Disentangled Latent Representation,
ACL 2019 Style Transformer: Unpaired Text Style Transfer without Disentangled Latent Representation, ACL 2019 -
Multiple Text Style Transfer by using Word-level Conditional Generative Adversarial
Network with Two-Phase Training Multiple Text Style Transfer by using Word-level Conditional Generative Adversarial Network with Two-Phase Training, EMNLP 2019 - Reconstruction(1), Back-Translation(2), Adversarial(3), Discriminator(4) Loss ܳ ݽف ࢎ ਊೞח ୨ - Generator ח GRU Decoder ੋؘ, Generation inputਵ۽ Style Condition ਸ ֍যח ߑध ࢎਊ - Reconstruction, Transfer Phase ܳ ܻ࠙೧ࢲ ण - Style Attribute ܳ Representation ਵ۽ ࢎਊೞӝ ٸ ޙী Multi-Attribute ਸ ਗೣ -
Multiple Text Style Transfer by using Word-level Conditional Generative Adversarial
Network with Two-Phase Training Multiple Text Style Transfer by using Word-level Conditional Generative Adversarial Network with Two-Phase Training, EMNLP 2019 -
- ؘఠח Sentiment Binary о ࠗ࠙, Multi-Attribute ա Dialog ؘఠо
աয়ח ୶ࣁ - ಣо ӝળ Style Intensity, Content Preservation, Naturalness о ݫੋ ܙ - Disentangled ࠁ Entangled ߑध જ. ݅ Entangled द णदఃӝח ࡆࣁ. - ਃ્ ࠗ࠙ ֤ޙٜ Multi-Attribute ח ӝࠄ ২࣌ - image style transfer ীࢲ ࢎਊ೮؍ GANਸ ӝ߈ਵ۽ ೞח ߑߨٜ ఫझীب ਊغ۰Ҋ ೞҊ . - ԙ Fine-Grained Attribute о ইפ؊ۄب Style Transfer ח оמೞ. ݅ ࠙ೠ ࢿמ ࠁغ ঋח. ݃ޖਵܻ ਃড -
хࢎפ✌ ୶о ޙ ژח ҾӘೠ ݶ ઁٚ ইې োۅ۽
োۅ ࣁਃ! ӣળࢿ (ML Research Scientist) [email protected] Facebook. @pingpong