Reviews are essential means of knowing the performance of a product.
In this project, I have created a model that predicts the score of a review based on the text.
This sentiment analysis model classifies the text into 1 to 5, based on the sentiment behind the review.
For example, "Nice product" usually means a score of 5 and “Poor quality” usually means a score of 1.
The model was trained using the
Amazon food reviews dataset,
which contains around 5 lakh reviews. Since there was a class imbalance, I did undersampling to balance the classes.
I used the BERT model and a linear layer at the end. Therefore, for word embedding, I used the BERT tokenizer.
The parameters of the BERT model were frozen during the training process to avoid computational complexity.
The test accuracy turned out to be 47.4%, much greater than the random case (20%).
