Style Transfer from Non-Parallel Text by Cross-Alignment Shen et al - - PowerPoint PPT Presentation

▶
style transfer from non parallel text by cross alignment
SMART_READER_LITE
LIVE PREVIEW

Style Transfer from Non-Parallel Text by Cross-Alignment Shen et al - - PowerPoint PPT Presentation

Jan 05, 2024 1.21k likes •1.43k views

Style Transfer from Non-Parallel Text by Cross-Alignment Shen et al 2017 Arxiv: 1705.09655 Presented by Leon Yin ML 2 Reading Group 2017-10-31 Maintain content and change style? View a sentence (x) of some distribution function of of style (y)

slide-1
SLIDE 1

Style Transfer from Non-Parallel Text by Cross-Alignment

Shen et al 2017 Arxiv: 1705.09655

Presented by Leon Yin ML2 Reading Group 2017-10-31

slide-2
SLIDE 2

Maintain content and change style?

View a sentence (x) of some distribution function of of style (y) and content (z). Style is sentiment between positive yelp reviews (3+ reviews) and negative. The two datasets are assumed to be talking about the same restaurants.

slide-3
SLIDE 3

Taco is z

x1 x2 These tacos are cold! y1= :( These tacos are the bomb! y2= :)

slide-4
SLIDE 4

Variational Auto-Encoder (VAE)

Image coutesty of: http://kvfrans.com/variational-autoencoders-explained/

slide-5
SLIDE 5

Pros and Cons of VAE?

“The fact that VAEs basically optimize likelihood while GANs

  • ptimize something else can be viewed both as an advantage
  • r a disadvantage for either one.”
  • Yoshua Bengio via Quora
slide-6
SLIDE 6

Two step solution

Encoder infers content (z) given sentence (x) and style (y). Generator returns sentence (x’) given style (y) from latent rep for content (z). This system can be trained using a GAN!

slide-7
SLIDE 7

Pros and Cons of VAE?

“The fact that VAEs basically optimize likelihood while GANs

  • ptimize something else can be viewed both as an advantage
  • r a disadvantage for either one.”
  • Yoshua Bengio via Quora
slide-8
SLIDE 8

Professor Forcing (Lamb et al 2016)

slide-9
SLIDE 9

Cross-Aligned Auto-Encoder (Shen et al 2017)

slide-10
SLIDE 10

Evaluation

Used pre-trained sentiment classifier with a prediction accuracy of 85.4%.

slide-11
SLIDE 11
slide-12
SLIDE 12

Taco is z?

x1 x2 These tacos are cold! y1= :( These tacos are the bomb! y2= :)

slide-13
SLIDE 13

What is z?

x1 x2 These tacos are cold! y1= :( This spaghetti is sooo Italian! y2= :)

slide-14
SLIDE 14

Open Questions

Is sentiment a good example of style? Other training systems like Professor Forcing? Emerging methods of evaluating and comparing GANs? How much time do you spend picking or exploring the data you feed into a model?

slide-15
SLIDE 15

Thanks!

“Translation is a matter of compromises.”

  • Ken Liu Reddit AMA
slide-16
SLIDE 16

Extra Slides For Questions...

slide-17
SLIDE 17

Data set for Pos X1 (n=250k) and Neg X2 (n=350k)

2 datasets w/ same content distro (Yelp reviews) and styles y1 (pos) and y2 (neg).

  • 3+ star reviews == positive.
  • Filter out reviews if

○ +10 sentences ○ +15 words / sentence.

Used to estimate the style transfer functions between X1 and X2 p(x1|x2;y1,y2) and p(x2|x1;y1,y2).

slide-18
SLIDE 18

Reconstruction Loss

slide-19
SLIDE 19