Attention based Neural Networks for Chemical Protein Relation

• Task description: http://www.biocreative.org/tasks/biocreative-vi/track-5/
• Data (may require login): http://www.biocreative.org/accounts/login/?next=/resources/corpora/chemprot-corpus-biocreative-vi/

Prerequisites

• Python 2.7
• TensorFlow 1.2.1
• Keras 2.0.5
• NLTK
• Scikit-learn
• ConfigParser

Usage

Update configuration file

Go through the config file config/main_config.ini to modify the following paths accordingly.

• corpus_dir: unzipped corpus directory
• out_dir: the output directory of preprocessed file

Preprocessing

python extract_sentences.py

By default, you will see the relation instances of train, dev and test sets under out_dir: training.txt, development.txt and test.txt.

Encode Sentences

Load word embeddings and generate word index (word2id) for the corpus by running:

python preprocess.py

A subset of word embeddings, vocabulary, and sentencse are stored under the compressed pkl file pkl/bioc_rel_ent_candidate.pkl.gz

Train Models

Load the encoded sentences, initalize model parameters, compile Tensorflow and Keras models, and run training and testing on the dataset by:

python dnn.py

The output file is in Brat standoff format, the same as the gold standard files. Each epoch of ATT-GRU will take 83 seconds on a NVIDIA Tesla P40 GPU to complete.

The offical envaluation script can be downloaded at the offical site: ChemProt evaluation kit A copy is also provided under ./eval for the convenience of usage, and it will be called automatically after the output file on the test set is generented.

The official result is:

Total annotations: 3458
Total predictions: 2939
TP: 1687
FN: 1771
FP: 1252
Precision: 0.5740047635250085
Recall: 0.48785425101214575
F-score: 0.5274347350320463

The confusion matrix and classification report will also be printed.

Confusion Matrix:

[[8859  138  573   70   73  272]
 [ 344  212   37    1    1    3]
 [ 485   33  969    3   11   11]
 [  69    4    1   86    0    1]
 [ 125    0    5    4  136    0]
 [ 274    3   12    0    0  280]]

Classification Report:

             precision    recall  f1-score   support

      CPR:3      0.544     0.355     0.429       598
      CPR:4      0.607     0.641     0.623      1512
      CPR:5      0.524     0.534     0.529       161
      CPR:6      0.615     0.504     0.554       270
      CPR:9      0.494     0.492     0.493       569

avg / total      0.570     0.541     0.551      3110

Visualization of Attention

Please follow the Jupyter Notebook model_att_vis.ipynb for details.

Examples:

Acknowledgment

The relation classification architechture is based on the Relation CNN implementation from UKPLab.

The attention RNN is inspired by the code snippet from cbaziotis.

Reference

S Liu, F Shen, R Komandur Elayavilli, Y Wang, M Rastegar-Mojarad, V Chaudhary, H Liu. Extracting chemical–protein relations using attention-based neural networks. Database, Volume 2018, bay102.