Initializing Convolutional Filters with Semantic Features for Text Classification

Source code of the EMNLP 2017 paper Initializing Convolutional Filters with Semantic Features for Text Classification. http://aclweb.org/anthology/D17-1201

The purpose of the project

The project implements the weight initialization method in the paper.

Recently, CNN [1] is proven to be effective for sentence classification.

Our method further improves the CNN baseline significantly with little additional costs.

One can easily obtain the state-of-the-art or competitive results via running demo.sh in our project.

Requirements and preprocessing

Code is written in Python3 and requires Tensorflow (>=1.0.0).

The data preprocessing and hyper-parameter setting strictly follow the implementation in https://github.com/yoonkim/CNN_sentence [1], whose preprocessing code is reused in our project.

Pre-trained embedding 'GoogleNews-vectors-negative300.bin' is available at https://drive.google.com/file/d/0B7XkCwpI5KDYNlNUTTlSS21pQmM/edit?usp=sharing.

How to use

To process the raw data:

python process_data.py vectors_path

This will create a pickle object called 'mr.p' in the same folder, which contains the dataset in the right format.

To obtain the pre-trained features for weight initialization:

python initialize_filter.py [--ngram NGRAM] mr.p

This will create a pickle object called 'weights_NGRAM.pkl', containing parameters for initializing weights in the CNN filters.

To train the model:

python cnn.py mr.p weights_NGRAM.pkl

This will train and test the model.

Reference

Yoon Kim. Convolutional neural networks for sentence classification. EMNLP 2014.

Shen Li et al. Initializing Convolutional Filters with Semantic Features for Text Classification. EMNLP 2017.