style transfer from non-parallel text by...

12
Style Transfer from Non-Parallel Text by Cross-Alignment Tianxiao Shen, Tao Lei, Regina Barzilay, Tommi Jaakkola NIPS 2017

Upload: others

Post on 10-Jul-2020

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Style Transfer from Non-Parallel Text by Cross-Alignmentink-ron.usc.edu/xiangren/ml4know19spring/presentations/Nazgol.pdf · Style Transfer from Non-Parallel Text by Cross-Alignment

Style Transfer from Non-Parallel Text by Cross-Alignment

Tianxiao Shen, Tao Lei, Regina Barzilay, Tommi JaakkolaNIPS 2017

Page 2: Style Transfer from Non-Parallel Text by Cross-Alignmentink-ron.usc.edu/xiangren/ml4know19spring/presentations/Nazgol.pdf · Style Transfer from Non-Parallel Text by Cross-Alignment

STYLE TRANSFER ON TEXT

• brief/verbose

• colloquial/professional

• personal/impersonal

• polite/impolite

Page 3: Style Transfer from Non-Parallel Text by Cross-Alignmentink-ron.usc.edu/xiangren/ml4know19spring/presentations/Nazgol.pdf · Style Transfer from Non-Parallel Text by Cross-Alignment

NON-PARALLEL DATA

• Parallel :

• corresponding output for each input

• Usually not available

Page 4: Style Transfer from Non-Parallel Text by Cross-Alignmentink-ron.usc.edu/xiangren/ml4know19spring/presentations/Nazgol.pdf · Style Transfer from Non-Parallel Text by Cross-Alignment

STYLE TRANSFER ON IMAGES

• Has had a lot of success

• Cycle GANs and other models…

• Not applicable to text

• because of discreteness of natural language

Page 5: Style Transfer from Non-Parallel Text by Cross-Alignmentink-ron.usc.edu/xiangren/ml4know19spring/presentations/Nazgol.pdf · Style Transfer from Non-Parallel Text by Cross-Alignment

PROPOSED MODEL

X is generated from p(x|y, z)

**Important assumption: two datasets have to have the same content.

Page 6: Style Transfer from Non-Parallel Text by Cross-Alignmentink-ron.usc.edu/xiangren/ml4know19spring/presentations/Nazgol.pdf · Style Transfer from Non-Parallel Text by Cross-Alignment

PROPOSED MODEL

• proposition 1: In this generative framework x1 and x2’s joint distribution can be recovered from their marginals only if for any different y1 and y2, distribution p(x|y1) and p(x|y2) are different.

• If the distribution of z has a more complex distribution, such as Gaussian mixture, then affine transformations can be uniquely determined.

Page 7: Style Transfer from Non-Parallel Text by Cross-Alignmentink-ron.usc.edu/xiangren/ml4know19spring/presentations/Nazgol.pdf · Style Transfer from Non-Parallel Text by Cross-Alignment

METHOD (ELEMENTARY)

• Encoder-Decoder

• Reconstruction loss

• Variational Auto Encoder (VAE)

• imposes prior density p(z), z ~ N(0, I)

• KL-divergence regularizer to align posteriors

Page 8: Style Transfer from Non-Parallel Text by Cross-Alignmentink-ron.usc.edu/xiangren/ml4know19spring/presentations/Nazgol.pdf · Style Transfer from Non-Parallel Text by Cross-Alignment

METHOD(ALIGNED AUTO-ENCODER)

• Relax the prior assumption on p(z)

• Use Lagrangian relaxation

• Adversarial loss

• Final loss:

Page 9: Style Transfer from Non-Parallel Text by Cross-Alignmentink-ron.usc.edu/xiangren/ml4know19spring/presentations/Nazgol.pdf · Style Transfer from Non-Parallel Text by Cross-Alignment

METHOD(CROSS-ALIGNED AUTO-ENCODER)

Page 10: Style Transfer from Non-Parallel Text by Cross-Alignmentink-ron.usc.edu/xiangren/ml4know19spring/presentations/Nazgol.pdf · Style Transfer from Non-Parallel Text by Cross-Alignment

EVALUATION

• Sentiment Modification

• Sentiment Accuracy

• Human Evaluation

• Word Substitution Decipherment

• Blue scores

• Word Order Recovery

• Blue scores

Page 11: Style Transfer from Non-Parallel Text by Cross-Alignmentink-ron.usc.edu/xiangren/ml4know19spring/presentations/Nazgol.pdf · Style Transfer from Non-Parallel Text by Cross-Alignment

SENTIMENT MODIFICATION

Page 12: Style Transfer from Non-Parallel Text by Cross-Alignmentink-ron.usc.edu/xiangren/ml4know19spring/presentations/Nazgol.pdf · Style Transfer from Non-Parallel Text by Cross-Alignment

EVALUATION