Simple text classification using DAN, LSTM and BERT for SMP-19-ECISA contest.
- python 3.7
- pytorch
- transformers
- pytorch-ignite
- numpy
Already uploaded the clean version of raw text, named "train.txt", "dev.txt" and "test.txt" for DAN and LSTM, and "bert_train.txt" and "bert_dev.txt" for BERT.
To training BERT, personally, I strongly recommend you to use the gpu version code, see "bert.py" instead of "bert.ipythonotebook".
SMP only offers label for train and dev, so I only report performance over dev set here, the metric here is accuracy.
Model | DAN | LSTM | BERT |
---|---|---|---|
Accuracy | 0.752 | 0.720 | 0.819 |