ORCID

https://orcid.org/0000-0002-3107-8914

Keywords

Natural language processing (NLP), Large language models (LLMs), Computational social science, Explainable artificial intelligence (XAI), Generative AI (GenAI), User experience research

Abstract

As digital platforms increasingly mediate political discourse, they have also become fertile ground for radicalized ideologies and targeted threats against public officials. These dynamics are intensified by the unchecked spread of extremist content and mounting regulatory pressure on platforms. This dissertation addresses this complex sociotechnical challenge by advancing the field of Artificial Intelligence (AI) for Social Wellness, developing scalable, explainable, and ethically grounded tools to detect ideological polarization and threats in online texts.

The first part focuses on ideological discourse analysis, examining how political bias and radicalization manifest in online communities. We introduce RICo (Reddit Ideology Classification), a dataset of 1.52M news articles from 55 subreddits labeled as liberal, conservative, or restricted (extremist). RICo supports binary and three-class classification, achieving 86.19% and 79.23% accuracy respectively. Beyond benchmarks, it provides a computational lens into language shifts from mainstream to radicalized worldviews.

The second part addresses threats in user-generated content from extreme clusters. We compiled 2.3M Telegram replies from 17 radical channels and proposed a three-tier taxonomy: no threat, judicial threat, and non-judicial threat, optimized for operational clarity. ALERT (Active Learning and Explainable AI for Risk-based Threat detection) reduces labeling effort by 86.5% via active learning and achieves over 90% across classification metrics with a domain-tuned RoBERTa+ model. ALERT integrates multiple explanation modalities to improve moderator transparency.

Building on ALERT’s explainability design, we developed an improved RoBERTa model that achieved 95.8% accuracy. The final part evaluates its real-world impact through TRuST-M (Threat Reasoning and User Study of Trust in Moderation). This within-subjects study found that clearer explanations significantly increased user trust, and that trust strongly correlated with perceived moderation effectiveness.

Together, this dissertation contributes novel algorithms, datasets, and frameworks for ethical AI deployment in content moderation, and news recommendation equipping researchers, practitioners, and policymakers to responsibly counter online extremism.

Completion Date

2025

Semester

Fall

Committee Chair

Jiann-Shiun Yuan

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

Department of Electrical and Computer Engineering

Format

PDF

Release Date

12-15-2026

Document Type

Dissertation

Campus Location

Orlando (Main) Campus

Subjects

Artificial intelligence--Social aspects; Artificial intelligence--Moral and ethical aspects; Terrorism in social media; Radicalization--Computer network resources; Discourse analysis--Political aspects

Download

Available for download on Tuesday, December 15, 2026

Share

COinS
 

Accessibility Statement

This item was created or digitized prior to April 24, 2027, or is a reproduction of legacy media created before that date. It is preserved in its original, unmodified state specifically for research, reference, or historical recordkeeping. In accordance with the ADA Title II Final Rule, the University Libraries provides accessible versions of archival materials upon request. To request an accommodation for this item, please submit an accessibility request form.