Tools
Tools
Sajikumar, Sreelakshmi (2024) Emotion Detection in Text: A Comprehensive Analysis Using Classical, Deep Learning, and Transformer-Based Models. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (1MB) | Preview |
Preview |
PDF (Configuration Manual)
Download (1MB) | Preview |
Abstract
The sentiment analysis of comments on social media is a challenging subproblem in Natural Language Processing with usage in mental health care, customer feedback, and political analysis. This study investigates the effectiveness of classical machine learning, deep learning, and transformer-based models in classifying text into six emotion categories: Positive emotions as wellbeing, happiness, love, and approving satisfaction, and negative as including pain, sorrow, rage, terror, shock, and loathing. Some of the concerns that the research meets include: Text noise, data skewness, and large-scale computational inferences.
The methodology employs a multi-phase approach: Logistic Regression & SVM using BoW and word embeddings as well as GloVe & Word2Vec were used to set the first benchmark. Long Short-Term Memory (LSTM) networks, especially the use of Bayesian LSTM, is performed owing to its contextual modeling approach. At last, several the state-of-the-art transformer models including fine-tuned BERT, RoBERTa models and Flan-T5 for few-shot learning, are used for better context understanding.
The study concludes that transformer-based models are more accurate, precise, recall and F1-score than other methods, but consume more computational power. To translate the theories into practice and make the findings of the study accessible, both Streamlit and Gradio-based web applications were created. The Streamlit app offers an intuitive interface for detecting and visualizing emotions in real-time, while the Gradio interface allows for quick deployment and testing of emotion detection models via a shareable link. This research offers tangible solutions on how to deal with noisy text, achieve an appropriate treatment of class imbalance, and enhance the performance of emotion detection systems, to improve theoretical knowledge, and advance practical applications.
Actions (login required)
 |
View Item |