Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Text Style Transfer Overview

Text Style Transfer Overview

Scatter Lab Inc.

January 23, 2020

Transcript

32 slides
  1. 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
  2. Style Transfer ঌইࠁӝ ޙઁ ੿੄ ޙ੢ীࢲ झఋੌ੉ۄח Ѫ਷ ޖ঺ੌөਃ? ৘द)

    • ޙ੢੄ ӛ੿, ࠗ੿ • ޙয୓, ҳয୓, ࢿ҃୓(?) • যൃ ࢶఖ (য۰਍ ਊয, ए਍ ਊয) • ݈ై (যܲझ۞਍ ݈ై, গੋՙܻ੄ ݈ై)
  3. 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 ޙ੢ਸ ౠ੿ झఋੌਸ ನೣೠ ޙ੢ਵ۽ ߸҃ೞח ޙઁੑפ׮. ױ, ߸҃दী ޙ੢੄ ஶబஎ ژח ޙ୓җ ҙ۲ হח ੿ࠁٜਸ ࠁઓ೧ঠ ೤פ׮ • ৘) ӛ੿੸ੋ ܻ࠭ -> ࠗ੿੸ੋ ܻ࠭ Text Style Transfer ۆ ޖ঺ੌөਃ?
  4. 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੉ ֫਷ ૑಴ܳ ࢸ੿ೞӝ য۰਑. - Content Preservation: ޙ੢੄ झఋੌ੉ ߄Շݶࢲ ӝઓ੄ ஶబஎо ߸ഋغয ߡܻח ޙઁо ੓਺ - ੿੄੄ ݽഐࢿ : झఋੌ੉ۄח Ѫਸ ݺഛೞѱ ੿੄ೞӝ য۵ӝ ٸޙী, ಽযঠೡ ޙઁо ݽഐೣ ੉ పझ௼ীࢲ ೧Ѿ೧ঠ ೡ ޙઁ
  5. 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)
  6. Style Transfer ঌইࠁӝ FYELP (Multi-Attribute) YELP ؘ੉ఠࣇী Gender, Category Attribute

    ܳ ୶оೠ ؘ੉ఠࣇ (MULTIPLE-ATTRIBUTE TEXT REWRITING, ICLR 2019)
  7. Style Transfer ঌইࠁӝ Empathetic Dialog Dataset (Multi-Attribute) ೧׼ ؀ച੄ х੿੉

    ۨ੉࠶݂ غয੓ח ؘ੉ఠࣇ. х੿ -> ׮ܲ х੿ਵ۽ ߸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)
  8. Style Transfer ঌইࠁӝ Transfer ػ ޙ੢੉ ݾ಴۽ ೞח झఋੌী ঴݃ա

    Ӕ੽ೠо? ఫझ౟੄ झఋੌਸ ߄Բח ӝࠄ੸ੋ ݾ಴о ׳ࢿ غ঻ח૑ ഛੋೞח ૑಴ Style Transfer Intensity ಣо ߑߨ - Style Classification (৘: Sentiment Classification) - Style Representation Similarity ৘: Ӓ ৔ച ѐԵ੨! (੉ ޙ੢੉ pos ੋо?) <- (input) Ӓ ৔ച ૓૞ ֢੨੉঻য, (target style) pos
  9. Style Transfer ঌইࠁӝ ӝઓ੄ ޙ੢ীࢲ ׸Ҋ ੓؍ ஶబஎܳ ୽࠙൤ ࠁઓ೮חо?

    Style Transferܳ ೞݶࢲ ޙ੢੄ ஶబஎо ࢚ࣚغѢա ߸ഋ غ঻ח૑ܳ ഛੋೞח ૑಴ Content Preservation (=Relevance) ಣо ߑߨ - BLEU, METEOR, Overlap - Representation Similarity - Sentence Matching Model ৘: Ӓ ৔ച ѐԵ੨! (੉ ޙ੢੉ inputҗ زੌೠ ஶబஎ?) <- (input) Ӓ ৔ച ૓૞ ֢੨੉঻য,
  10. Style Transfer ঌইࠁӝ Naturalness (=Fluency) ࢤࢿػ ੉ ޙ੢੉ ੗োझ۞਍ ޙ੢ੋо?

    যࣽ੉ ੜޅغযب content preservation, style transfer intensity ח ֫ਸ ࣻ ੓਺ ಣо ߑߨ - Language Model Perplexity - Adversarial Evaluation (੉ѱ ࢎۈ੉ ॵ ޙ੢ੌө ࠈ੉ ॵ ޙ੢ੌө?) ৘: Ե੨! Ӓ ৔ച, ޷ଢ଼ (੉ ޙ੢ ઁ؀۽ ࢤࢿ ⪟פ?) <- (input) Ӓ ৔ച ૓૞ ֢੨੉঻য,
  11. Style Transfer ঌইࠁӝ Tradeoff Evaluating Style Transfer for Text, 2019

    NAACL Style Transfer Intensity ৬ Content Preservation, Naturalness ח Tradeoffо ࢤӡ ࣻ ߆ী হ਺
  12. Disentanglement о ަө? VAE Latent Representation Content Style Disentangled VAE

    Latent Variable ਋ܻח Style ਸ Control ೞҊ र਷ؘ, Latent Representation ੉ Content ৬ Ԙৈ੓׮… ੉ ࢚క۽ ޥо Latent Representationܳ ߸ഋೠ׮ݶ contents ب ߸ഋؼ૑ ށۄ! ഒઙ?! Content ৬ Style р੄ Ү଱ਸ ୭ࣗചೞҊ, Style ݅ ઑ੘ೞ੗!
  13. 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
  14. Disentangled Representation Learning for Non-Parallel Text Style Transfer, ACL 2019

    Disentangled Representation Learning for Non-Parallel Text Style Transfer, ACL 2019
  15. 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?!
  16. MULTIPLE-ATTRIBUTE TEXT REWRITING MULTIPLE-ATTRIBUTE TEXT REWRITING, ICLR 2019 - Style

    Latent Representation ਸ ܻ࠙ೞח ߑߨ (Disentanglement Method)੉ ਬബೞ૑ ޅೞ׮ח Ѫ ਸ प೷ਸ ా೧ ୊਺ ૐݺೠ ֤ޙ - Denoising auto-encoding ਸ ࢎਊ೧ࢲ Style Tokenਸ Masking ೞҊ ੉ܳ Style Condition੉ ઱য઎ਸ ٸ ೧׼ ష௾ਸ ݏ୶ח ޙઁ. - Style Token Classification ੉ ೙ਃ۽ೣ - Multiple Attribute ޙઁܳ ಽ঻਺ 
 х੿ -> ׮ܲ х੿, ৈࢿ -> թࢿ, ੻਷ -> ֢ੋ - ֤ޙী Figure о হ׮….
  17. 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 о੿ೞҊ झఋੌ੉ ߸҃ػ ޙ੢ਸ ଺ইࠁ੗! - ߸ചػ Style੄ Latent Variable ਸ ଺ӝ ਤ೧ Predictor ੄ Error Surface ীࢲ Gradient Discent ܳ ࢎਊ. ী۞о ୭ࣗച غח ನੋ౟о ਋ܻо ଺Ҋ੗ ೞח ޙ੢੉ۄҊ о੿ೞҊ ޙઁܳ ಽӝ - Predictor = VAE + Attribute Predictor + Content Preservation ೧Ѿਸ ਤ೧ CBOW بੑ
  18. Revision in Continuous Space: Unsupervised Text Style Transfer without Adversarial

    Learning Revision in Continuous Space: Fine-Grained Control of Text Style Transfer, AAAI 2019
  19. MULTIPLE-ATTRIBUTE TEXT REWRITING - AAAI 2019 ֤ޙҗ Entangled Latent Repre

    ܳ Gradient ߑधਵ۽ ࣻ੿ೠ׮ח ੽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
  20. 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
  21. Style Transformer: Unpaired Text Style Transfer without Disentangled Latent Representation,

    ACL 2019 Style Transformer: Unpaired Text Style Transfer 
 without Disentangled Latent Representation, ACL 2019
  22. 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 ֍য઱ח ߑध ࢎਊ - Reconstruction, Transfer Phase ܳ ܻ࠙೧ࢲ ೟ण - Style Attribute ܳ Representation ਵ۽ ࢎਊೞӝ ٸ ޙী Multi-Attribute ਸ ૑ਗೣ
  23. 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
  24. - ؘ੉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੉ ੼੼ ఫझ౟ীب ੸ਊغ۰Ҋ ೞҊ ੓׮. - ԙ Fine-Grained Attribute о ইפ؊ۄب Style Transfer ח оמೞ׮. ׮݅ ୽࠙ೠ ࢿמ੉ ࠁ੢غ૓ ঋח׮. ݃ޖਵܻ ਃড
  25. хࢎ೤פ׮✌ ୶о ૕ޙ ژח ҾӘೠ ੼੉ ੓׮ݶ ঱ઁٚ ইې োۅ୊۽

    োۅ ઱ࣁਃ! ӣળࢿ (ML Research Scientist) [email protected] Facebook. @pingpong

More Decks by Scatter Lab Inc.

Other Decks in Research