text-classification

Turkish Text Classification task

For text classification task, we trained 3 different models:

• Bi-directional Long Short Term Memory
• CNN-Text Model
• BERT: Pretrained Bi-directional Transformers

Environment

• torch==1.8.1
• transformers==4.11.3

Download datasets

Download and extract data from the v1.0 release of mukayese-datasets:

mkdir -p datasets
cd datasets/
mv /your/download/path/offenseval.zip .
mv /your/download/path/news-cat.zip .
unzip offenseval.zip
unzip news-cat.zip

Dataset statistics

	Offenseval	News-cat
Avg. #words	8.5	227.3
#Classes	2	5
Training	28,000	750
Validation	3277	150
Test	3515	250
Total	34792	1150

Running the task

python -m src.main

Baselines expected performance

	offenseval	news-cat	Avg.
BiLSTM	0.747	0.808	0.777
CNN Text	0.751	0.883	0.817
BERT	0.823	0.944	0.883