Event_Extraction

A biomedical event extraction system by a novel Combination Strategy based on hybrid deep neural networks.
Online demo : http://www.predictor.xin/event_extraction/
Zhu L, Zheng H. Biomedical event extraction with a novel combination strategy based on hybrid deep neural networks[J]. BMC bioinformatics, 2020, 21(1): 47. https://rdcu.be/b1Pn1

Uasge

Preprocess

(0. The train corpus contains .txt/.a1/.a2 files, placed in /txt /a1 /a2 directories respectively.)

Run corpus_table.py, convert .txt/.a1/.a2 files into uniform .csv files (saved in /table).
Run a2_preprocess.py, convert .a2 files into structured .csv files (saved in /a2_table).
Run a2_preprocess_modification.py, convert event modifications in .a2 files into structured .csv files (saved in /a2_modification_table).
The test corpus should be generated by repeating above process and saved in others directories, e.g. /table_test,/a2_table_test,/a2_modification_table_test.

Training

Run my_word2vec.py to generate the pre-trained word vectors (better to modify the code to use a large external corpus).
Run pretrain_char_cnn.py to generate the pre-trained character-level CNN.
Run train_net.py to train the main neural networks, in format python train_net.py NUMBER with a argument NUMBER which number the trained networks, the trained networks are saved in /nets.

Testing

Run test_generate_strategy_best_ensemble.py to extract events from testset in /table_test, in format python test_generate_strategy_best_ensemble.py NUM1 NUM2 where NUM1 denotes the number of ensemble epoches and NUM2 denotes the number of ensemble training. Meanwhile, the base number for NUM1 and NUM2 should be set in codes. The result .a2 files is saved in /a2_result. (1. The test_generate_event_eval_ensemble.py is a variation method with same usage.)
Use python BioNLP-evaluation-CG.py -r /GOLDEN_STANDARD_DIR /PARENT_DIR/a2_result/*.a2 to evaluate the performance (for CG).

Requirements

Pytorch
Gensim
NLTK
Pandas
sklearn