Chung, Master Ek-Karat (2024) Developing a fine-tuned transformer model to detect social media hate speech texts. Final Year Project, UTAR.
Abstract
This research explores the development and evaluation of a hate speech detection system using transformer-based models, focusing on the robustness, efficiency, and scalability of the model. The study emphasizes key design considerations, including scalability, which addresses the model's capability to handle large volumes of data, and accuracy, achieved through fine-tuning methods for transformer models like BERT. Reviewing model to do proper performance analysis on existing model in detecting Social Media Hate Speech Texts such as Long Short-Term Memory (LSTM) and Bidirectional Gated Recurrent Unit (Bi-GRU), GigaBERT for Arabic Hate Speech Detection, BERT-Based Approaches, DistilBERT and RoBERTa, T5 and Electra, Comparison of Transformer Models and its Challenges and Limitation. Besides, it also briefly discusses on system design to ensure that the model is conceptually accurate, scalable, and maintainable, providing a flexible framework for ongoing research in hate speech detection on social media. The research also discusses on the facing challenges such as data imbalance, computational limitations, and extensive hyperparameter tuning, all of which were addressed through various techniques and strategies. This research show system's experiment/ simulation to show performance with evaluated using a Logistic Regression model on a split dataset, fine-tuning with GridSearchCV, how the model's accuracy improved. The experiment successfully show a predictive model with high accuracy and precision, also indicated future improvements in detecting hate speech on social media. The results underscore the importance of ongoing refinement in machine learning models or deep learning model to address complex, real-world issues such as hate speech detection.
Actions (login required)
![[img]](http://eprints.utar.edu.my/style/images/fileicons/application_pdf.png)
