Project Text Sentiment Classification

Abstract:

The aim of this project is to classify tweets into a 'happy' and a 'sad' class. To fulfill that aim, we implemented different machine learning algorithms and finally choose a Recurrent Neural Network as the best method. We achieved an accuracy of about 86% which is close to the state of the art.

Requirements:

• Python 3
• Tensorflow
• Keras
• ScikitLearn

How to reproduce:

1. Download the datasets on CrowdAI
2. Put them into a directory called data/
3. Download twitter GloVe embeddings over here
4. Put them into a directory called data/glove.twitter.27B
5. Create an empty directory called output/ and put into it the trained model you downloaded over here
6. Run the file run.py with python 3. It should produce a file prediction.csv that you can upload directly on CrowdAI

Content:

• run.py, a simple script that load the saved model and predict the classes of the test dataset
• mode_selection.py, a simple script that we used to cross-validate the accuracy of different classification algorithms
• rnn.py, the implementation of our Recurrent Neural Network
• plot.py, different functions that we used to make the plot of the report