Covid-19 sentiment analysis by fine-tuning bert model basesd on social network.
Also, this repo provides and e2e demo 👨‍💻

• Flask
• Pytorch
• Transformers

[Optional] Create and activate your virtual environments first!

$ conda create -n venv
$ source activate venv

Begin to fine-tune

$ python main.py

When training complete, main.py will save accuracy and loss history(training/validate), then we provide two functions in predict.py to get inference (get_predictions / get_prediction_with_single). This step will infernece from test dataloader using function get_predictions, where live-demo will inference from web's input in live-demo using function get_prediction_with_single.

We use flask as web server, and interact with web client via ajax. Due to file size limit, we place fine-tuned model weight here, of course, you can change to your weight under folder web_demo

$ cd web_demo
$ python web.py

We would like to add bertviz at first, but it seems unsupport on html for now.

• https://www.kaggle.com/surajkum1198/datasets

• https://www.kaggle.com/alankritamishra/covid-19-tweet-sentiment-analysis
• https://leemeng.tw/attack_on_bert_transfer_learning_in_nlp.html#1.-%E6%BA%96%E5%82%99%E5%8E%9F%E5%A7%8B%E6%96%87%E6%9C%AC%E6%95%B8%E6%93%9A