This model will uitilize the dataset that have multiple labels. It will have n+1-heads according to n-tasks and a MaskedLM head.
Our method achieves amazing result with our NEU, VSFC and ViHSD datasets (before we add layernorm hehe):
| Task | Accuracy | F1 macro | F1 weighted |
|---|---|---|---|
| NEU sentiment | 84.42 | 85.15 | 84.43 |
| NEU classification | 81.33 | 73.98 | 81.57 |
| VSFC sentiment | 93.94 | 83.77 | 94.19 |
| VSFC topic | 89.45 | 80.82 | 90.15 |
| ViHSD | 88.31 | 68.49 | 88.77 |
To train the model, modify the model config in train.py and run
python3 train.py
We made a website for the implementation of the model, you can checkout here
If you are seeing this, it means that we havent finished documenting our code. Please be patient