Date of Graduation
Summer 2026
Degree
Master of Science in Computer Science
Department
Computer Science
Committee Chair
Jamil Saquer
Abstract
The widespread proliferation of hate speech on social media platforms poses significant challenges for content moderation and user safety, requiring automated systems that are simultaneously accurate, efficient, and capable of fine-grained distinctions. This thesis investigates hate speech detection through five published manuscripts organized into two complementary threads: binary detection (hateful vs. non-hateful) and multiclass detection across demographic targeting categories. The binary thread progresses from a broad 38-model baseline spanning traditional machine learning, deep learning, and transformer architectures (where RoBERTa reaches 91.48% accuracy and CatBoost remains competitive at 88.60%) to parameter-efficient adaptation, in which Low-Rank Adaptation (LoRA) of large language models attains state-of-the-art performance: Phi-4 reaches 91.84% accuracy and 92.79% F1-score while training only 0.15% of its parameters, and zero-shot API-based classification offers an immediate-deployment alternative at roughly nine percentage points lower accuracy. The thread culminates in BERT-OTA, an ontology-guided dual-stream architecture combining transformer attention with a graph convolutional network that reaches 91.30% accuracy. The multiclass thread evaluates 303 model configurations across the same three paradigms on 39,747 tweets spanning five demographic categories, with RoBERTa-large reaching 95.06% accuracy, and then introduces RoBERTa-OTA, which integrates ontological knowledge to reach 96.04% accuracy and 96.06% F1-score, surpassing the RoBERTa baseline and prior state of the art, with its largest gains on the most challenging gender-based and other-hate categories while adding only 0.33% parameter overhead. Across all studies, raw text consistently outperforms aggressive preprocessing for transformer models, and parameter-efficient and ontology-guided techniques deliver strong accuracy at modest computational cost, enabling deployment across diverse computational environments.
Keywords
hate speech detection, transformer models, large language models, parameter-efficient fine-tuning, ontology integration, graph neural networks, multiclass classification, zero-shot classification, content moderation, natural language processing
Subject Categories
Artificial Intelligence and Robotics | Data Science
Copyright
© Mahmoud Abusaqer
Recommended Citation
Abusaqer, Mahmoud, "Advancing Hate Speech Detection: Binary and Multiclass Approaches From Traditional Methods to Parameter-Efficient and Ontology-Guided Language Models" (2026). Graduate Theses/Dissertations. 4207.
https://bearworks.missouristate.edu/theses/4207
