BERT-fine-tuning-for-twitter-sentiment-analysis

• Download the pre-trained models

  download the pre-trained models from the https://github.com/google-research/bert#pre-trained-models , unzip the .zip file and put the files in the Bert_base_dir.

• Run the run_classifier.py

  cd the dir test_bert
  python ./run_classifier.py
  --task_name=twitter
  --do_train=true
  --do_eval=true
  --data_dir=../data
  --vocab_file=../Bert_base_dir/vocab.txt
  --bert_config_file=../Bert_base_dir/bert_config.json
  --init_checkpoint=../Bert_base_dir/bert_model.ckpt
  --max_seq_length=64
  --train_batch_size=32
  --learning_rate=2e-5
  --num_train_epoch=3.0
  --output_dir=../model
  
  the output model will in the model dir

• Prediction

  python ./run_classifier.py
  --task_name=twitter
  --do_predict=true
  --data_dir=../data
  --vocab_file=../Bert_base_dir/vocab.txt
  --bert_config_file=../Bert_base_dir/bert_config.json
  --init_checkpoint=../model
  --max_seq_length=64
  --output_dir=../data/bert_result
  
  the prediction result will in the bert_result dir,if you want to test the acc,you can handle it by youself.In the /data/bert_result/test_result.tsv ,the first column is the probability of class 0.

note! I Run the model in win10,so there is some different.