higgsboson-x / bert-sentiment-analysis

Sentiment analysis using pretrained Bert model

Bert-Sentiment-Analysis

Sentiment analysis using pretrained Bert model

Dependencies

python3.6
torch==1.3.1
torchtext==0.5.0
transformers==2.8.0

Prepare Data

Prepare test.json and train.json in the data directory. Each file contain lines of json strings with following fields.

{
  "text": [string, the sentence],
  "label": [integer, true sentiment label],
  "id": [integer, sentence id]
}

Training

Run the following command.

$ cd code
$ python train.py \
    --data_path [data path] \
    --epochs 40 \
    --epochs_per_val 5 \
    --save_id [id] \
    --device_id [device id] \
    --bidirectional

Here are some more options.

optional arguments:
  -h, --help            show this help message and exit
  --data_path DATA_PATH
                        data path, contains `train.json` and `test.json`
  --epochs EPOCHS       training epochs
  --epochs_per_val EPOCHS_PER_VAL
                        epochs per evaluation
  --batch_size BATCH_SIZE
                        batch size of data
  --device_id DEVICE_ID
                        gpu index
  --save_id SAVE_ID     model save id
  --h_dim H_DIM         hidden dimension of GRU
  --n_cls N_CLS         number of classes
  --n_layers N_LAYERS   number of layers
  --bidirectional       whether to use bidirectional GRU
  --dropout DROPOUT     dropout rate

Testing

Run the following command.

$ cd code
$ python eval.py --data_path [data path] --save_id [id] --device_id [device id]