This work was a final project for a course on neural networks during my master's degree. It is based on a CodaLab competition for performing sentiment analysis on data from Twitter. The project was designed to test students on knowledge of machine/deep learning, not only the practical implementation of a network using Keras but also our capacity for troubleshooting and domain knowledge.

This was an especially challenging task because there are 11 categories for classification of each tweet, corresponding to emotions e.g. anger, joy, sadness. With 11 categories, a model must perform very well to breach even the halfway mark for accuracy. For example, an accuracy of 30% on the test set was the baseline performance needed in order to receive a B grade on the project, and 45% for an A. I experimented with several different architectures, including a basic feedforward network and a convolutional neural network (CNN). I ended up getting my best performance (49% accuracy) with an LSTM after determining the most effective architecture and choice of optimizer and loss function.

This repository contains the necessary twitter data to train and test the model. However, because of Github's space constraints, I am not able to upload the GLoVe vectors I used to this repository. The pre-trained vectors can be easily downloaded here: https://nlp.stanford.edu/projects/glove/.