TensorFlow implementation of the paper Adversarial Multi-task Learning for Text Classification.

• The code uses CNN instead of LSTM.
• The Gradient Reversal Layer is copied from https://github.com/pumpikano/tf-dann.
• The Orthogonality Constraints loss (diff loss) is copied from 'research/domain_adaptation' of https://github.com/tensorflow/models. The correlation matrix is normalized. Otherwise, the loss value will be too large.

performace using 50d word embedding:

To train the model:

cd data/
tar zxvf fudan-mtl-dataset.tar.gz
cd ../
python3 src/main.py --build_data
python3 src/main.py --adv