Deep learning project on summarizing multiple Amazon reviews of a product into keywords and presenting them in a word cloud.
Authors are ordered alphabetically by last name.
- Yuying Fan (@fyy26)
- Rehaan Furniturewala (@r3khaan)
- Andrew Raine (@andrewlr09)
Even though online shoppers have thousands of reviews to learn about each product, they lack the time to read each review and make the optimal purchasing decision.
We used a dataset of Amazon Reviews from a research group at the University of California San Diego. The dataset that we are able to access, a “small” subset of the full set, contains 75,257,650 reviews that we can train on.
Each review is of the format: (overall_score, verified_purchase, review_time, reviewer_id, product_id, style, reviewer_name, review_text, summary).
We used a BERT tokenizer across all our models for generating word embeddings.
- Build a classification model that predicts the star-rating (1-5) of a review based on its text.
- Run part-of-speech tagging on each review using the NLTK library.
- For each adjective in a given review, replace the adjective with BERT's
[MASK]token to produce a masked review. - Run each masked review through the star-rating prediction model and compute the difference in the predicted star-rating between the original and masked review.
- Collect the few adjectives with the biggest difference to generate a summary word cloud.
Random forest with 100 estimators; validation accuracy of 20.93% after training on 90% of the data (no better than a random classifier).
forest = sklearn.ensemble.RandomForestClassifier(n_estimators=100)
Stacked RNN with 5 layers and a hidden size of 50, followed by a linear output layer; best validation accuracy of 39.65%.
Built on top of a pre-trained BERT; validation accuracy in the low 60% after removing bias.
