Naive Bayes vs BERT on Emotion Classification

Description: This project compares the performance of a Naive Bayes model and a BERT model on emotion classification from text.

• Using a emotion classification dataset from huggingface (https://huggingface.co/datasets/dair-ai/emotion), the models are trained and evaluated on the dataset.
• a comprehensive grid search for Naive Bayes is performed to find the best hyperparameters which resulted in an accuracy of 0.8385.
• the pre-trained BERT model is used in an out-of-the-box or fine-tuned (just head or full model) approach with the best one achieving an accuracy of 0.9320.

report.pdf contains a description of all our experiments and results.

Installation

Before running the project, you need to set up the required environment. Follow these steps:

1. Clone the Repository:

git clone https://github.com/jantiegges/emotion-classification-from-text.git
cd emotion-classification-from-text

2. Create a Virtual Environment (Optional but Recommended):

python -m venv venv
source venv/bin/activate  # On Windows, use `venv\Scripts\activate`

3. Install Dependencies:

pip install -r requirements.txt

Usage

To use this project, follow these steps:

1. Run Jupyter Notebooks:

• Launch Jupyter Notebook in the project directory:

jupyter notebook

• Open the relevant Jupyter notebooks, such as:
  • experiments.ipynb - contains all of the experiments
  • data_analysis.ipynb - contains data analysis

2. Explore the Code:

• Review the codebase:
  • models/ - contains the naive bayes and BERT models
  • utils/ - contains data preprocessing functions
  • main.py - contains functions for running grid search and experiments

3. Customize and Experiment:

• Feel free to customize parameters and experiment with the code.
• Note any additional instructions provided within the notebooks.