README
Offensive Language Identification for Dravidian languages
This code is for the "Syntax-aware Offensive Content Identification in Low-resourced Code-mixed Languages with Continual Pre-training" paper.
You can use the model with the parameters in the train.json file. Syntax-BERT model is adopted from"Improving BERT with Syntax-aware Local Attention"
Usage
python main.py train.json
model options (code in train.json)
- bert: BERT
- cont-BERT: continual training BERT
- syntax_bert: Syntax-BERT
To train Cont-Syntax-BERT, you need to run cont-BERT than run syntax-BERT with the trained model of cont-BERT.
Results
Results on DravidianCodeMix dataset:
Tamil
| Model | Precision | Recall | F1-Score |
|---|---|---|---|
| BERT | 0.53 | 0.73 | 0.61 |
| Syntax-BERT | 0.55 | 0.73 | 0.62 |
| Cont-BERT | 0.75 | 0.77 | 0.76 |
| Cont-Syntax-BERT | 0.84 | 0.76 | 0.80 |
Kannada
| Model | Precision | Recall | F1-Score |
|---|---|---|---|
| BERT | 0.50 | 0.64 | 0.56 |
| Syntax-BERT | 0.71 | 0.76 | 0.73 |
| Cont-BERT | 0.72 | 0.74 | 0.73 |
| Cont-Syntax-BERT | 0.77 | 0.76 | 0.76 |
Malayalam
| Model | Precision | Recall | F1-Score |
|---|---|---|---|
| BERT | 0.83 | 0.88 | 0.78 |
| Syntax-BERT | 0.87 | 0.90 | 0.87 |
| Cont-BERT | 0.94 | 0.95 | 0.94 |
| Cont-Syntax-BERT | 0.96 | 0.97 | 0.96 |
Contributors
5 commits
Similar repos
Languages
Python
100.0%
README
Offensive Language Identification for Dravidian languages
This code is for the "Syntax-aware Offensive Content Identification in Low-resourced Code-mixed Languages with Continual Pre-training" paper.
You can use the model with the parameters in the train.json file. Syntax-BERT model is adopted from"Improving BERT with Syntax-aware Local Attention"
Usage
python main.py train.json
model options (code in train.json)
- bert: BERT
- cont-BERT: continual training BERT
- syntax_bert: Syntax-BERT
To train Cont-Syntax-BERT, you need to run cont-BERT than run syntax-BERT with the trained model of cont-BERT.
Results
Results on DravidianCodeMix dataset:
Tamil
| Model | Precision | Recall | F1-Score |
|---|---|---|---|
| BERT | 0.53 | 0.73 | 0.61 |
| Syntax-BERT | 0.55 | 0.73 | 0.62 |
| Cont-BERT | 0.75 | 0.77 | 0.76 |
| Cont-Syntax-BERT | 0.84 | 0.76 | 0.80 |
Kannada
| Model | Precision | Recall | F1-Score |
|---|---|---|---|
| BERT | 0.50 | 0.64 | 0.56 |
| Syntax-BERT | 0.71 | 0.76 | 0.73 |
| Cont-BERT | 0.72 | 0.74 | 0.73 |
| Cont-Syntax-BERT | 0.77 | 0.76 | 0.76 |
Malayalam
| Model | Precision | Recall | F1-Score |
|---|---|---|---|
| BERT | 0.83 | 0.88 | 0.78 |
| Syntax-BERT | 0.87 | 0.90 | 0.87 |
| Cont-BERT | 0.94 | 0.95 | 0.94 |
| Cont-Syntax-BERT | 0.96 | 0.97 | 0.96 |
Similar repos
Contributors
5 commits
Languages
Python
100.0%