Movie Review Sentiment Classifier

- By Subhalingam D

The model has been trained using the IMDB dataset obtained from Kaggle. Pretrained word embeddings from GloVe were used with 2 LSTM layers to predict 0 for negative review and 1 for a positive one.

The whole model was trained in Python (using Notebook in Colab). Keras was used primarily with Tensorflow backend.

Prediction

It's straightforward to use predict.py to predict with your own review, unless you catch an error message!

Requirments

The basic dependencies (after installing python) are numpy, keras with tensorflow backend. The same can be installed using the following Terminal command:

pip install -r requirements.txt

Make sure you have model.h5 and tokenizer.json in the same directory as predict.py. You can get all these files from my repo by clicking here

Running the python code

predict.py requires one mandatory and accepts one optional command-line argument. You can run the file by the following template command:

python predict.py -r "<REVIEW>" [-o]

• -r should be followed by the review within double quotes
• -o if present, will display the prediction value (between 0 and 1) at the end

For example,

python predict.py -r "excellent story and what a finishing! so many twists and should i even say about the direction?"

python predict.py -r "it's a good movie from the story standpoint but the screenplay was pathetic" -o

How to interpret the predicted value when -o flag is used?

This is quite straightforward. The predicted value will be between 0 and 1 which will be rounded to the nearest integer. After that, 0 would mean a negative review whereas 1 would mean a positive review.

So if the predicted value is 0.009686, it is very close to 0 and signifies that the predictor strongly believes that it is a negative review. However, if the predicted value is 0.574038, it would mean positive but that's an intermediate value.

In other words, this value would be a measure of confidence of the prediction-the extreme ends would mean a more strong prediction.

---

Note: This was an individual project done for learning purpose and several blogs were referred in the process. I would like to mention this blog specifically, which I used to learn how to preprocess data (removing punctuations and tokenising)

For any queries, contact Subhalingam at subhalingam.d@gmail.com