GA-Reader

Code accompanying the paper Gated Attention Reader for Text Comprehension.

Prerequisites

• Python 2.7
• Theano (tested on 0.9.0dev1.dev-RELEASE) and all dependencies
• Lasagne (tested on 0.2.dev1)
• Numpy (>=1.12)
• Maybe more, just use pip install if you get an error

Preprocessed Data

You can get the preprocessed data files from here. Extract the tar files to the data/ directory. Ensure that the symbolic links point to folders with training/, validation/ and test/ directories for each dataset.

You can also get the pretrained Glove vectors from the above link. Place this file in the data/ directory as well.

To run

Issue the command:

python run.py --dataset <wdw|cnn|dailymail|cbtcn|cbtne>

Complete list of options:

$ python run.py --help
Using gpu device 0: GeForce GTX TITAN X (CNMeM is disabled, cuDNN 5105)
usage: run.py [-h] [--mode MODE] [--nlayers NLAYERS] [--dataset DATASET]
              [--seed SEED] [--gating_fn GATING_FN]

optional arguments:
  -h, --help            show this help message and exit
  --mode MODE           run mode - (0-train+test, 1-train only, 2-test only,
                        3-val only) (default: 0)
  --nlayers NLAYERS     Number of reader layers (default: 3)
  --dataset DATASET     Dataset - (cnn || dailymail || cbtcn || cbtne || wdw)
                        (default: wdw)
  --seed SEED           Seed for different experiments with same settings
                        (default: 1)
  --gating_fn GATING_FN
                        Gating function (T.mul || Tsum || Tconcat) (default:
                        T.mul)

To set dataset specific hyperparameters modify config.py.

Note

Make sure to add THEANO_FLAGS=device=cpu,floatX=float32 before any command if you are running on a CPU.

Contributors

If you use this code please cite the following:

Dhingra, B., Liu, H., Yang, Z., Cohen, W. W., & Salakhutdinov, R. (2016). Gated-Attention Readers for Text Comprehension. arXiv preprint arXiv:1606.01549.

@article{dhingra2016gated,
  title={Gated-Attention Readers for Text Comprehension},
  author={Dhingra, Bhuwan and Liu, Hanxiao and Yang, Zhilin, and Cohen, William W and Salakhutdinov, Ruslan},
  journal={arXiv preprint arXiv:1606.01549},
  year={2016}
}

Report bugs and missing info to bdhingraATandrewDOTcmuDOTedu (replace AT, DOT appropriately).