Segment CNN for relation classification

Requirements

Code is written in Python (2.7) and requires Theano (0.9).

Data Preprocessing

To process the raw data, run

python data.py 

This is a wrapper code calling cnn_preprocess.embed_train_test() with arguments specifying word embedding dimensions (e.g., 200) and padding length (e.g., 7). This will create a pickle object (e.g., semrel_pp200_pad7.p) in the directory 'data/semrel_pp', which contains the dataset with the right components to be used by cnn_semrel.py.

Using the GPU

GPU will result in a good 10x to 20x speed-up compared to CPU, so it is highly recommended. For example:

THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32,gcc.cxxflags=-march=corei7 python cnn_semrel.py -static -word2vec -img_w200 -l1_nhu100 -pad7 -trp

Example output

GPU output (mif is micro-averaged f-measure):

epoch: 1, training time: 4.86 secs, train perf: 70.54 %, mif: 55.50 %
epoch: 2, training time: 4.70 secs, train perf: 76.67 %, mif: 61.89 %
epoch: 3, training time: 4.75 secs, train perf: 78.77 %, mif: 64.75 %

To get embeddings we used

https://northwestern.box.com/s/eprxyxmee37p3d6khqbpn125tyttq4u6

Include embeddings trained on MIMIC3 corpus, dimensions from 100 to 600.

Citation

@article{luo2017segment,
  title={Segment convolutional neural networks (Seg-CNNs) for classifying relations in clinical notes},
  author={Luo, Yuan and Cheng, Yu and Uzuner, {\"O}zlem and Szolovits, Peter and Starren, Justin},
  journal={Journal of the American Medical Informatics Association},
  volume={25},
  number={1},
  pages={93--98},
  year={2017},
  publisher={Oxford University Press}
}