Perform information extraction on several japanese documents. Each document has 20 tags, as shown in figures below.  The experiment includes 82 training data and 22 development data.
python3.6 preprocessing.py --train_path {path_to_train_data} --dev_path {path_to_dev_data} --test_path {path_to_test_data}
python3.6 train.py --backbone {bert, cnn, dualbert}
python3.6 predict.py --model {path_to_model} --mode {dev, test} --backbone {bert, cnn, dualbert}or simply
python3.6 test.py {path_to_test_set_directory}| Model | Precision | Recall | F1 score |
|---|---|---|---|
| BERT(char only) | Text | Text | 95.82 |
| BERT+CNN | 96.34 | 96.38 | 96.16 |
| BERT(char+word) | 96.11 | 96.51 | 96.10 |
r08922115 吳冠霖
r08944034 洪商荃
r08944035 呂翊愷