Emotion Detection: A Comparative Analysis

A comprehensive analysis of Machine Learning and Transformer models for multi-label emotion detection on the GoEmotions dataset.

This repository contains the code and resources for the project "Emotion Detection: A Comparative Analysis," which evaluates the performance of classic machine learning models against modern Transformer architectures. The key achievement of this project is a fine-tuned RoBERTa model that demonstrates high sensitivity with a state-of-the-art recall score.

🚀 Features

• Comprehensive Analysis: Compares 4 different models (Logistic Regression, Random Forest, DistilBERT, and RoBERTa).
• Multi-Label Classification: Tackles the complex task of predicting multiple emotions for a single piece of text.
• In-Depth EDA: Includes detailed Exploratory Data Analysis on the GoEmotions dataset, highlighting the severe class imbalance.
• High-Recall Model: The fine-tuned RoBERTa model achieves a weighted average recall of 0.66, demonstrating high sensitivity.
• Code: All code is provided in easy-to-follow Jupyter/Colab notebooks.

📖 Methodology

The project workflow is divided into two parallel approaches:

1. Classic Machine Learning Baseline:

   • Text is vectorized using TF-IDF.
   • Logistic Regression and Random Forest models are trained using a MultiOutputClassifier.

2. Transformer Fine-Tuning:

   • Text is tokenized using specific tokenizers for DistilBERT and RoBERTa.
   • The pre-trained models are fine-tuned on the GoEmotions dataset using PyTorch and Hugging Face.

📊 Results

The results clearly show the superiority of Transformer models. Our fine-tuned RoBERTa model achieved the best performance, most notably a high recall score, indicating its effectiveness at identifying emotions.

A key finding was the precision-recall trade-off, a direct consequence of the dataset's class imbalance. While our model excels at finding emotions (high recall), it sometimes over-predicts, leading to lower precision.

🛠️ How to Use

To run this project, follow these steps:

1. Clone the repository:

2. Install the dependencies:

   pip install -r requirements.txt

3. Download the dataset:
   The GoEmotions.csv dataset kaggle or Google's github.

4. Run the notebooks:
   Open the files in the notebooks/ directory using Jupyter Notebook, JupyterLab, or Google Colab.

👥 Authors

• Mahmud - GitHub - LinkedIn
• Tanjila Hussen