COMP_5222_Project

This is the repository of COMP 5222 group project, our group number is 13.

How to run:

0. Download all files in this repository.

1. Prepare the train and dev data:

Open your terminal, navigate to the "code" folder and run DataProcess.py.

! python DataProcess.py --data_path ../data

2. Train the model (need a >=10 GB GPU server):

Open your terminal and run flair_train.py in the "code" folder.
Here you can change the model structure flexibly. (refer to the tutorial of flair).

! python flair_train.py --input ../data \
              --output ../output \
              --gpu 'cuda' \
              --train_file 'fold_1234.txt' \
              --dev_file 'fold_5_dev.txt' \
              --test_file 'fold_5_test.txt' \
              --transformer 'bert-base-uncased' \
              --learning_rate 2e-5 \
              --mini_batch_size 8 \
              --max_epochs 20 \
              --patience 2

3. Make prediction and compute F1 score:

Open your terminal and run predict.py in the "code" folder. (you need to modify data_path according to your situation)
---------predict_file: the path of file that your model predicts.
---------sample_pred_file: the path where you want to save the standard prediction.

! python predict.py --input ../data \
              --output ../output \
              --gpu 'cuda' \
              --train_file 'fold_1234.txt' \
              --dev_file 'fold_5_test.txt' \
              --test_file 'fold_5_test.txt' \
              --checkpoint 'best-model.pt' \
              --predict_file 'bert_fold_5_test.txt' \
              --sample_pred_file 'fold_5_test_recover.txt'

Then you can get a result like this:
F1 score: 0.6704417291328144