Author: Zeping Yu
This work is accepted by COLING 2018.
Sliced Recurrent Neural Network (SRNN).
SRNN is able to get much faster speed than standard RNN by slicing the sequences into many subsequences.
The code is written in keras, using tensorflow backend. We implement the SRNN(8,2) here, and Yelp 2013 dataset is used.
keras version: 2.1.5
tensorflow version: 1.6.0
python : 2.7
If you have any question, please contact me at [email protected].
The pre-trained GloVe word embeddings could be downloaded at:
https://nlp.stanford.edu/projects/glove/
The Yelp 2013, 2014 and 2015 datasets are at:
https://figshare.com/articles/Yelp_2013/6292142
https://figshare.com/articles/Untitled_Item/6292253
https://figshare.com/articles/Yelp_2015/6292334
Yelp_P, Amazon_P and Amazon_F datasets are at: https://drive.google.com/drive/folders/0Bz8a_Dbh9Qhbfll6bVpmNUtUcFdjYmF2SEpmZUZUcVNiMUw1TWN6RDV3a0JHT3kxLVhVR2M
Here is an interesting modification of SRNN for text generation, similar to language model: https://github.com/whackashoe/srnn/tree/text-generation