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:
| models | avg error |
|---|---|
| mtl | 13.75 |
| mtl + adv | 12.79 |
| mtl + adv + diff | 12.70 |
To train the model:
cd data/
tar zxvf fudan-mtl-dataset.tar.gz
cd ../
python3 src/main.py --word_dim=50 --build_data
python3 src/main.py --word_dim=50 --adv
python3 src/main.py --word_dim=50 --adv --test