The reviews are extracted from review.json in Yelp Dataset by using json_to_csv_converter.py.
130 K reviews are processed and stored into (yelp_review_first_130K_with_sentiment.csv) with review_dataset_extraction.ipynb.
For each review, if stars >= 4, we call its sentiment positive; if stars <= 2, we call its sentiment negative. The data are split into train (100 K), validation (10 K), and test (20 K).
Models used for sentiment analysis are LSTM (and Bidirectional LSTM) trained from scratch, and fine-tuned DistilBERT from HuggingFace Transformers.
Run the LSTM model rnn.ipynb in Jupyter Notebook;
Run the DistilBERT model with: python bert.py.