Research Overview

The overarching goal of this research program is to build trustworthy language intelligence for security-relevant and harmful online text. The lab develops reusable methods and assets for (i) incident-centric discourse triage and prioritization, (ii) auditable dataset curation and quality measurement, and (iii) robust and reproducible evaluation of traditional ML, transformers, and LLM-based approaches, including Green NLP perspectives (accuracy–latency–energy trade-offs).

Research Themes

1) Cybersecurity & Social-Media Incident Intelligence

Social platforms contain real-time signals during cyber events but are noisy and uneven in quality. This line of work focuses on incident-centric filtering, grading, and prioritization of posts that mention a specific cyberattack (e.g., healthcare incidents), using transparent features, reproducible pipelines, and analyst-friendly outputs.

  • CyberTweetGrader&Labeler (CTGL): domain-specific NLP pipeline for prioritizing cyber-incident discourse on Twitter/X.
  • Healthcare case studies: time-aware collection and analysis around health-system cyberattacks.

CTGL project page

2) Harmful, Abusive, and AI-Generated Text

This theme studies harmful language phenomena—cyberbullying and abusive language—and the growing challenge of distinguishing AI-generated from human-written text. The emphasis is on rigorous evaluation and reporting of error modes, dataset considerations, and deployment-relevant trade-offs.

  • Cyberbullying detection: comparative analysis across traditional ML and transformer models on Twitter/X.
  • AI vs. human text detection: empirical evaluation using lexical and syntactic baselines and controlled settings.

3) Benchmarking, Reliability, and Green NLP

This line focuses on controlled comparisons between model families—TF-IDF baselines, fine-tuned encoders, and prompted LLM inference—under explicit constraints. The goal is to characterize accuracy, robustness, throughput/latency, cost, and energy per inference, and to provide practical guidance for selecting models for real deployments.

  • ML vs. transformers vs. LLMs: reliability- and cost-aware evaluation protocols.
  • Green NLP: accuracy–latency–energy trade-offs for online abuse detection.

How to Collaborate

Collaboration is welcome on datasets, evaluation assets, and applied deployments in cybersecurity and safety-relevant text. If you are interested, please email and include (1) your background, (2) your target venue or timeline, and (3) whether you prefer a systems, NLP, or evaluation contribution.

© 2026 Muhammad Abusaqer