Sentiment Prediction on Large Movie Review Dataset

The goal of this project is to understand the sentiment from a given text.

This project uses the Large Movie Review Dataset (also known as IMDB dataset) for sentiment analysis. The dataset contains movie reveiws - every review has been labelled as either positive or negative. The problem is to correctly predict if a review is positie or negative. For more details about the dataset, please see: http://ai.stanford.edu/~amaas/data/sentiment/.

To make the problem (slightly) harder, we only read 250 words from every review. The average review size is around 300 words and some commonly reffered examples use first 500 words from every review.

Network Architecture

First, we need to convert the words to vectors using an embedding layer. Since every word is originally represented as an integer, we need to convert them to vectors of real numbers so that vectors of similar words are "closer". You can read more about embedding layer in Keras here: https://keras.io/layers/embeddings/.

Key Features

This network is not designed for the best accuracy. Instead, the goal is to make an efficient network which can do well while using minimun bumber of parameters. One way to quantify the score that includes test accuracy and the number of paramaters of the network is:

This network gives a test accuracy of over 88% (88.376% in my last test). Since there are 322,046 total parameters in the network, we get score = 0.2486 which is not too bad.

Network Summary
Total number of parameters: 322,046
Test Accuracy: 88%
Training Accuracy: 91%
Convolutional Layers: 3
Training Time: 14 seconds per epoch

Secret Sauce

Yes, it is not a secret anymore! The key decision that makes the number of parameters small, without compromising the accuracy, is to reduce the output size of embedding layer. Typical value of 132-dimensional output vector generates millions of parameters in just the embedding layer! While I have reduced the number of feature maps in the convolutional layers and the fully connected layer, the major improvement comes from reducing the dimension of vectors by the embedding layer.

Accuracy vs Epochs

Loss vs Epochs

Requirements

1. Python 3.5 (not tested with other versions)
2. Keras 2+ (Tested with TensorFlow Backend - might work with Theano but not tested)
3. Matplotlib for plotting convergence curves

Running the Code

The standalone code is in the file sentiment_CNN.py. If the required packages are configured properly, the code should run without any problem.

Disclaimers and Acknowledgements

I have taken the basic syntax from various existing projects on the web.