viix-co/cnn-rnf

Convolutional Neural Networks with Recurrent Neural Filters

Convolutional Neural Networks with Recurrent Neural Filters

Author: Yi Yang

Contact: [email protected]

Basic description

This is the Python implementation of the recurrent neural filters for convolutional neural networks, described in

Yi Yang
"Convolutional Neural Networks with Recurrent Neural Filters"
EMNLP 2018

[pdf]

BibTeX

@inproceedings{yang2018convolutional,
  title={Convolutional Neural Networks with Recurrent Neural Filters},
  author={Yang, Yi},
  booktitle={Proceedings of Empirical Methods in Natural Language Processing},
  year={2018}
}

Dependencies

1. TensorFlow
2. Keras
3. Optional: CUDA Toolkit for GPU programming.

Data

We use the Stanford Sentiment Treebank (SST) datasets processed by Lei et al. (2015). Please put all the files of this directory into the data/sst_text_convnet folder.

Please download the pre-trained GloVe vectors and unzip it into the data folder.

Results

Running the code requires two steps:

1. Prepare the data and generate the required data files

   # binary sentiment classification
   python proc_data.py data/stsa.binary.pkl
   
   # fine-grained sentiment classification
   python proc_data.py --train-path data/sst_text_convnet/stsa.fine.phrases.train \
                       --dev-path   data/sst_text_convnet/stsa.fine.dev \
                       --test-path  data/sst_text_convnet/stsa.fine.test \
                       data/stsa.fine.pkl
   

2. CNNs for sentiment classification with linear filters and recurrent neural filters (RNFs)

   # binary sentiment classification
   python cnn_keras.py --filter-type linear data/stsa.binary.pkl
   python cnn_keras.py --filter-type rnf data/stsa.binary.pkl
   
   # fine-grained sentiment classification
   python cnn_keras.py --filter-type linear data/stsa.fine.pkl
   python cnn_keras.py --filter-type rnf data/stsa.fine.pkl
   

Hyperparameter tunning may be needed to achive the best results reported in the paper.

Unfortunately, I failed to find out how to entirely eliminate randomness for training Keras-based models. However, you should be easily able to achieve 89%+ and 52%+ accuracies with RNFs after a few runs.

Recurrent neural filters consistently outperform linear filters across different filter widths, by 3-4% accuracy.