CNN-DCNN text autoencoder

Implementations of the models in the paper "Deconvolutional Paragraph Representation Learning" by Yizhe Zhang, Dinghan Shen, Guoyin Wang, Zhe Gan, Ricardo Henao and Lawrence Carin, NIPS 2017

Prerequisite:

• CUDA, cudnn
• Tensorflow (version >1.0). We used tensorflow 1.2. Run: pip install -r requirements.txt to install requirements

Run

• Run: python demo.py for reconstruction task
• Run: python char_correction.py for character-level correction task
• Run: python semi_supervised.py for semi-supervised task
• Options: options can be made by changing option class in the demo.py code.

• opt.n_hidden: number of hidden units.
• opt.layer: number of CNN/DCNN layer [2,3,4].
• opt.lr: learning rate.
• opt.batch_size: number of batchsize.

• Training roughly takes 6-7 hours (around 10-20 epochs) (for recontruction task) to converge on a K80 GPU machine.
• See output.txt for a sample of screen output for reconstruction task.

Data:

• Download from :
  • Reconstruction: Hotel review (1.52GB)
  • Char-level correction: Yahoo! review (character-level, 451MB)
  • Semi-supervised classification: Yelp review (629MB)

Citation

Please cite our paper if it helps with your research

• Arxiv link: https://arxiv.org/abs/1708.04729

@inproceedings{zhang2017deconvolutional,
  title={Deconvolutional Paragraph Representation Learning},
  author={Zhang, Yizhe and Shen, Dinghan and Wang, Guoyin and Gan, Zhe and Henao, Ricardo and Carin, Lawrence},
  Booktitle={NIPS},
  year={2017}
}

For any question or suggestions, feel free to contact [email protected]