Reimplement of CasRel Using Torch
Paper "A Novel Cascade Binary Tagging Frameword for Relational Triple Extraction" ACL 2020
The Original code was written in keras
- transformers 3.1.0
- torch 1.5.1+cu101
- tqdm
Using WebNLG DataSet and I preprocess this dataset.
链接:https://pan.baidu.com/s/1RLgcRR1pRXCaBxR5NrA5AQ 密码:8nqn
This Data has been preprocessed
Get the pre-trained Bert Model
-
Download the bert-base-uncased
-
mkdir under data directory
mkdir data/static mv {Default-Path}/bert-base-uncased data/static/
Note : Please following above shell command, do not use your own path
Train the Model and Test
python src/run.pyThe pure model performance
| Acc | Precision | Recall | F1-score | |
|---|---|---|---|---|
| subject_start_point | 0.994 | 0.943 | 0.964 | 0.953 |
| subject_end_point | 0.994 | 0.946 | 0.962 | 0.954 |
| object_start_point | 0.999 | 0.923 | 0.766 | 0.837 |
| object_end_point | 0.999 | 0.927 | 0.764 | 0.838 |
Decoding Part is not finished. QAQ
You can check the logger.py under the folder of log for more detail training process.
| Precision | Recall | F1-score | |
|---|---|---|---|
| End2End | 0.662 | 0.517 | 0.577 |
In end2end test, model dosen't perform good as it descriped in Paper.
I am still find the problem
! ! !
Set epoches a max value, and have enough patient
In my Training Experiment, I set epoches = 100.
-
When epoch equals to 23, the subject tagger starts to recall
-
When epoch equal to 31, the subject tagger starts to recall.
-
Before this, in validation, the precision, recall and f1-score equal to zero.