An implementation for Relation Classification using textCNN
Most part of code refers to roomylee,a little modification have been made to adapt current version of tensorflow.
The model is descriped as below in Paper Relation Extraction: Perspective from Convolutional Neural Networks
To make the actual algorithm clear, I would like to present you this picture:
- Tensorflow-1.11
- scikit-learn
┃━ SemEval2010_task8_all_data
┃━ SemEval2010_task8_training
┃━ TRAIN_FILE.TXT
┃━ TRAIN_TEST_DISTRIB.TXT
┃━ runs
┃━ 1547716564(timestap)
┃━ checkpoints
┃━ summaries
┃━ summaries
┃━ train
┃━ dev
┃━ configure.py
┃━ data_helpers.py
┃━ model.py
┃━ configure.py
┃━ train.py
┃━ utils.py
┃━ GoogleNews-vectors-negative300.bin
- The train data is located in "SemEval2010_task8_all_data/SemEval2010_task8_training/TRAIN_FILE.TXT".
- "GoogleNews-vectors-negative300" is used as pre-trained word embeddings.Or you can get it at BaiduYun , 2muq
- The parameters setting is located in configure.py.You can get help by command
python train --help - To run :
python train (--[parameter_name] [value])
- Test data is located in "SemEval2010_task8_all_data/SemEval2010_task8_testing_keys/TEST_FILE_FULL.TXT".
- Parameter checkpoint_dir and the vacab_to_index,pos_to_index json file(generated by train step) paths text_tokenizer_path、pos_tokenizer_path must be given.Just like below:
python eval.py --checkpoint_dir "runs/1523902663/checkpoints/" --text_tokenizer_path "runs/1523902663/checkpoints/text_tokenizer.json --pos_tokenizer_path "runs/1523902663/checkpoints/pos_tokenizer.json"
- Use macro average F1 score and Accuracy to evaluate the performance.
- I got a F1 score of 74.85% and the accuracy is 76% by using default settings
- Used dropout and l2 regularization , but overfitting is still a big problem .
- The best performance may reach 82% for F1, you can adjust the parameters to get better performance .