Roy Xie

I am a first-year Computer Science Ph.D. student at Duke University, advised by Bhuwan Dhingra. My research focuses on the properties of language models from an adversarial perspective. My recent work involves developing attacks and evaluation strategies to study the robustness and limitations of language models.

My research is supported by the NSF Graduate Research Fellowship. Previously, I completed my undergraduate degree at George Mason University, where I worked with Antonios Anastasopoulos on multilingual NLP.

/ / / /

Research

My research interests span various aspects of natural language processing and machine learning, focusing on building safe, interpretable, and robust language technologies for social good.

	LLM-Resistant Math Word Problem Generation via Adversarial Attacks Roy Xie, Chengxuan Huang, Junlin Wang, Bhuwan Dhingra Preprint, 2024 A novel approach to generate math word problems that LLMs are unable to solve, while preserving the coherence and difficulty of the original problems.
	Raccon: Prompt Extraction Benchmark of LLM-Integrated Applications Junlin Wang, Tianyi Yang, Roy Xie, Bhuwan Dhingra Under review, 2024 A benchmark which comprehensively evaluates a LLM’s susceptibility to prompt extraction attacks.
	Tailoring Vaccine Messaging with Common-Ground Opinions Rickard Stureborg, Sanxing Chen, Roy Xie, Aayushi Kunjal Patel, Christopher Li, Chloe Zhu, Tingnan Hu, Jun Yang, Bhuwan Dhingra Findings NAACL, 2024 A comprehensive dataset for training and evaluating models for tailoring vaccine messaging to opinions to establish common ground.
	Extracting Lexical Features from Dialects via Interpretable Dialect Classifiers Roy Xie, Orevaoghene Ahia, Yulia Tsvetkov, Antonios Anastasopoulos NAACL, 2024 Extract lexical features from language dialects through interpretable dialect classifiers.
	GMNLP at SemEval-2023 Task 12: Sentiment Analysis with Phylogeny-Based Adapters Md Mahfuz Ibn Alam∗, Roy Xie∗, Fahim Faisal∗, Antonios Anastasopoulos SemEval@ACL, 2023 A sentiment analysis system for low-resource African languages, leveraging multilingual language models, data augmentation method, and phylogeny-based adapter-tuning.
	Noisy Parallel Data Alignment Roy Xie and Antonios Anastasopoulos Findings EACL, 2023 Make word-level alignment models more robust under OCR noisy setting by using noise simulation and structural bias.

Website template