The original dataset ai.stanford.edu/~amaas/data/sentiment/

The project consists of two parts. Neural Network classifier (sources in /classifier) and web-server on django.

File presentation.ipynb allows you to run all experiment on google colab.

I used 2 popular Neural net architectures: GRU and Conv net.

Nets predict review score from 1 to 10.

Both models trained on 3 epochs with Adam optimizer and negative log likehood loss.

There are results of several experiments. Although models predict score from 1 to 10, in the original task reviews with score more than 7 considered positive and less than 5 considered as negative. All columns except Real accuracy calculated according to this rule.

As we can see ConvNet has a better score. Thus, it is out choice for website.

All trained model parameters can be found here