We attempt to infer the ratings of food reviews based on sentiment analysis using Bidirectional LSTMs with PyTorch.

Dataset download from https://www.kaggle.com/snap/amazon-fine-food-reviews/kernels

• Full dataset has 500,000+ reviews
• Balanced sample has 147500 reviews

• Python 3.7
• Standard Anaconda packages (e.g., numpy, pandas)
• PyTorch 1.0.0
• Torchtext 0.3.1
• Spacy
• GloVe word embeddings: 6B 100d. https://nlp.stanford.edu/projects/glove/

• 70% accuracy on full dataset (60% on balanced test set)
• 0.69 F1 score
• 0.47 mean absolute error (in other words, on average, our predictions are only off by less than 0.5 stars)
• 0.36 seconds of training per epoch on one Nvidia Tesla K80. 5 - 10 epochs are required to reach convergence.

$ python predict.py "It was very fragrant and delicate. I definitely recommend." --print
Probability distribution:  [0.00227669 0.00171347 0.01063003 0.22481784 0.760562  ] 
5

$ python predict.py "I think it was enjoyable" --print
Probability distribution:  [0.05328422 0.11656603 0.20920141 0.4209052  0.20004311] 
4

$ python predict.py "It was slightly better than edible." --print
Probability distribution:  [0.1383879  0.23203094 0.30232716 0.23983285 0.08742111]  
3

$ python predict.py "It was disappointing." --print
Probability distribution:  [0.415574   0.4342737  0.14427686 0.00485574 0.00101966]  
2

$ python predict.py "It smelled foul and was very greasy." --print
Probability distribution:  [0.8074387  0.16070081 0.02730567 0.00314497 0.00140986] 
1

Once we download and extract Review.csv from Kaggle to the /data directory, we can use this command $ python train.py

And evaluate it on the test set using this command $ python eval.py

train.py automatically calls dataloader.py which prepares the data for balancing and loading.

FastText word embeddings