Predicting Emotions with GRU
Description
The goal of the project is see how machine learning can categorize the sentiment behind textual data. To do this, we trained a Gated Recurrent Unit (GRU) model on a labeled dataset of tweets.
Tasks
- Built and trained a single-layer GRU model with Keras/TensorFlow on a 70k-tweet dataset for six-class sentiment prediction, using tokenization and word embeddings.
- Tuned hyperparameters via learning curves, analyzed misclassifications with confusion matrices, and observed model validity by visualizing embedded vectors with color maps.
- Achieved 93% test accuracy with a 60k/10k train-test split, improving by 5% over other models.
Tools
