Sentiment Analysis with BERT

Basic structure

• model/ folder: all the models and its components.
• setup/ folder: the python and cuda requirements.
• utils/ folder: the data-handling and evaluation parts.

Datasets

The datasets will be automatically downloaded when running the code. There are 3 datasets available:

• Yelp13
• Yelp14
• IMDB

Setup

> sh setup/install_cuda.sh
> pip3 install -r setup/requirements.txt

How to run the models

To run the models do:

> python3 -m model.train.<train_model_script> <parameters>

Example:

> python3 -m model.train.train_vanilla_bert --epochs 12 --learning_rate 0.001 --gradient_accumulation 8 --output_dir ./training_output

Available paramaters can be looked up in model/train/train.py.

Outputs

Outputs are saved under the followinf file structure: <model_name>/<dataset_name>/

The followinf files are saved:

• pytorch_model.bin: the model
• args.json: arguments of the model
• results.json: training results