rsarxiv / rc-cnn-dailymail

CNN/Daily Mail Reading Comprehension Task

CNN/Daily Mail Reading Comprehension Task

Code and trained models for the paper:

A Thorough Examination of the CNN/Daily Mail Reading Comprehension Task.

Dependencies

• Python 2.7
• Theano >= 0.7
• Lasagne 0.2.dev1

Datasets

• The two processed RC datasets:

  • CNN: http://cs.stanford.edu/~danqi/data/cnn.tar.gz (546M)
  • Daily Mail: http://cs.stanford.edu/~danqi/data/dailymail.tar.gz (1.4G)

  The original datasets can be downloaded from https://github.com/deepmind/rc-data or http://cs.nyu.edu/~kcho/DMQA/. Our processed ones are just simply concatenation of all data instances and keeping document, question and answer only for our inputs.

• Word embeddings:

  • glove.6B.zip: http://nlp.stanford.edu/data/glove.6B.zip

Usage

Training

    THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32
    python main.py --train_file /u/nlp/data/deepmind-qa/cnn/train.txt
                   --dev_file /u/nlp/data/deepmind-qa/cnn/dev.txt
                   --embedding_file /u/nlp/data/deepmind-qa/word-embeddings/glove.6B.100d.txt

Hyper-prameters

• relabeling: default is True.
• hidden_size: default is 128.
• bidir: default is True.
• num_layers: default is 1.
• rnn_type: default is "gru".
• att_func: default is "bilinear".
• batch_size: default is 32.
• num_epoches: default is 100.
• eval_iter: default is 100.
• dropout_rate: default is 0.2.
• optimizer: default is "sgd".
• learning_rate: default is 0.1.
• grad_clipping: default is 10.

Reference

    @inproceedings{chen2016thorough,
        title={A Thorough Examination of the CNN/Daily Mail Reading Comprehension Task},
        author={Chen, Danqi and Bolton, Jason and Manning, Christopher D.},
        booktitle={Association for Computational Linguistics (ACL)},
        year={2016}
    }