NIPS2018_DECAPROP

Implementation of Densely Connected Attention Propagation for Reading Comprehension (NIPS 2018) - Yi Tay, Luu Anh Tuan, Siu Cheung Hui, Jian Su.

This model achieves quite competitive performance on four RC benchmarks (SearchQA, NewsQA, Quasar and NarrativeQA).

https://arxiv.org/abs/1811.04210

Model Code

The general idea here is that ./model/span_model.py contains the main span model and ./model/decaprop.py contains the DecaProp implementation. Bidirectional Attention Connectors (BAC) implementation is found at ./tylib/lib/att_op.py.

from tylib.lib.att_op import bidirectional_attention_connector
# c and q are sequences of bsz x seq_len x dim.
# seq_len may be different
# the output ff is the propagated feature.
c, q, ff = bidirectional_attention_connector(
                  c, q, c_len, q_len,
                  None, None,
                  mask_a=cmask, mask_b=qmask,
                  initializer=self.init,
                  factor=32, factor2=32,
                  name='bac')

Prep Scripts

You may find them at ./prep/ where datasets such as Squad, NewsQA, SearchQA and Quasar are found. Many of our pre-processing scripts reference https://github.com/nusnlp/amanda. Open domain QA dataset preprocessing were obtained from https://github.com/shuohangwang/mprc (reinforced reader ranker codebase by Wang et al.)

Please make a directory named ./corpus/ (for hosting raw datasets) and ./datasets/ for hosting prep-ed files. The key idea is that we prep the dataset into an env.gz file for training/evaluation.

Notes and Disclaimer

Most of the relevant code have been uploaded to this repository. I currently do not have the GPU resources to re-validate this repository. Assuming I didn't accidentally omit any code (while copying from my main repository and removing irrelevant/WIP code), this repository should run fine (the entry point is train_span.py, more running notes will be added when I have time).

The arguments in the argparser do not represent the optimal hyperparameters (from the time of NIPS'18 experiments, many other experiments were conducted, which may have changed the default hyperparameters). However, just couple of weeks ago I managed to get similar scores for searchqa/quasar.

Another useful note is that i use a language-based compositional control for model architecture, using if statements and keyword to control which the graph construction. This is controlled by --rnn_type in argparser. Also note that due to some tensorflow version upgrade issues, the cudnn CoVe LSTM is not working for the time being.

References

If you find our repository useful, please cite our paper:

@article{DBLP:journals/corr/abs-1811-04210,
  author    = {Yi Tay and
               Luu Anh Tuan and
               Siu Cheung Hui and
               Jian Su},
  title     = {Densely Connected Attention Propagation for Reading Comprehension},
  journal   = {CoRR},
  volume    = {abs/1811.04210},
  year      = {2018},
  url       = {http://arxiv.org/abs/1811.04210},
  archivePrefix = {arXiv},
  eprint    = {1811.04210},
  timestamp = {Fri, 23 Nov 2018 12:43:51 +0100},
  biburl    = {https://dblp.org/rec/bib/journals/corr/abs-1811-04210},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Acknowledgements

Several useful code bases we used in our work:

1. https://github.com/HKUST-KnowComp/R-Net (for cudnn RNNs and base R-NET model)
2. https://github.com/nusnlp/amanda (thanks for the evaluators and preprocessors which were useful!)
3. https://github.com/shuohangwang/mprc (For preprocessing of searchqa and quasar!)