Simple Convolutional Neural Network (CNN) for text classification at character level. In this project we will implement a Chinese Movie Sentiment (Positive/Negative) Classifier with CNN using TensorFlow. A manually labeled small corpus (HK movie reviews) has been uploaded for you to try out the network.
Mostly reused code from https://github.com/dennybritz/cnn-text-classification-tf which was posted by Denny Britz. Check out his great blog post on CNN classification http://www.wildml.com/2015/12/implementing-a-cnn-for-text-classification-in-tensorflow/.
Things are modified from the original author:
- modified the data helpers to support Chinese to train the embeddings at character level
- added Tensorboard visualization for embeddings
- added
sample.pywhich will generate prediction output - added
serve.pywhich will serve the classifier via Flask
- Tensorflow 1.0 or up
To train with the Chinese movie reviews dataset, run:
python train.py
To visualize results and embeddings, run:
tensorboard --logdir ./runs/1492654198/summaries/
To predict from a trained model, run with checkpoint_dir argument
python sample.py --checkpoint_dir=./runs/1492654198/checkpoints/ --text="套戲好鬼悶"
To serve the model via a Flask API, run:
python serve.py --checkpoint_dir=./runs/1492656039/checkpoints/
then send a HTTP POST request to http://localhost:5000/predict with the following json body:
{"text": "套戲好鬼悶"}
Some updates will be published soon.
- Train the embeddings at word level using jieba's tokenizer or other pre-trained word embeddings
- Use high level APIS such as
tf.layersortf.keras(will be available in TensorFLow 1.2)