nvn01234/pcnnwwla

PCNN with word-level attention

Requirements

• numpy

• tensorflow

• keras

Usage

Input file

Default input: input.txt

Input file can contain multiple line. Each line must be follow this pattern

relation e1start e1end e2start e2end word-1 word-2 ... word-n

• relation is integer. List of relations provided at origin_data/relations.txt

• e1start, e1end, e2start, e2end are starting index and ending index of two entities in the sentence

• From word-1 to word-n are the preprocessed sentence (tokenized, lowercase)

Example one line of input file:

0 0 0 4 4 lệ_quyên sinh ra tại hà_nội , trong gia_đình có 4 anh_chị_em , cô là con út

Train

Run initial.py one time to preprocess data. Preprocessed data will be placed in data folder

python initial.py

Run train.py to train model. You can edit model's parameters in settings.py. Train output and log will be placed in output/train/** folder. ** is the timestamp, you can train model many time to get best weights.

python train.py

Train output and log

Stored in output/train/**:

• result.json: macro-averaged f1-score, max f1-score, best fold, best weights

• fold_* folder: contain weights and all metrics of one fold

Test

Run test.py

python test.py example_test_input/weights.h5 example_test_input/input.txt

• example_test_input/weights.h5: path to weights of model

• example_test_input/input.txt: path to input file

Test output files

Stored in output/test/**:

• Predictions for input file: output.txt

• Error sentences: error_list.txt, error_predictions.txt