Building a simple regressor that predicts the review star of given food text review, 80% accuracy was achieved and about 0.38 Mean-Squared Error (MSE).
A neural network with Embedding layer as first layer, Long & Short Term Memory (LSTMs) since text is sequential data, then one fully connected neuron (dense) with linear activation function for regression ( continuous ratings ). The basic architecture is in the image below:
Amazon Fine Food Reviews: large dataset (more than 500K reviews ) that consists of reviews of fine foods from amazon.
Download and extract Reviews.csv to data folder (training will not work without)
- keras for creating, training and testing model.
- sklearn for utils like splitting data.
- tqdm for printing progress bars.
- pandas
- numpy
pip3 install requirements.txt
In case you want to test directly.
python test.py "Best Product Ever"
Output:
4.82/5
There is already a trained model in results folder. However you can tune some parameters in config.py to improve MSE such as number of LSTM units, embedding size, etc. then run:
python train.py