17 publications
2026
In Journal of Open Innovation: Technology, Market, and Complexity (Q1)
A hybrid Bayesian Belief Network and Artificial Neural Network approach for predicting Sustainable Development Goal outcomes from country-level AI readiness indicators, pairing interpretable dependency structure with non-linear predictive power.
In IEEE (Accepted)
This IEEE article provides empirical measurements of LLM-generated texts in multilingual disinformation, moving beyond speculation to analyze the growing presence and sophistication of AI-generated content across diverse linguistic contexts and cultural settings.
2024
In Computer
This Computer article examines how Generative AI disproportionately harms longtail users, focusing on the structural inequalities that emerge when AI systems are deployed without adequate consideration for vulnerable user populations with limited resources to combat disinformation.
In IEEE Intelligent Systems
This IEEE Intelligent Systems article examines the "longtail" impact of Generative AI on disinformation in high-impact events and resource-limited settings. We analyze four critical dimensions—quantity, quality, personalization, and hallucination—proposing strategies to combat disinformation risks in vulnerable communities and during critical events where consequences are most profound.
2026
In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP ‘26)
A benchmark for over-refusal in large audio language models — how often they decline pseudo-harmful but benign spoken queries, and what drives it.
In Proceedings of the 43rd International Conference on Machine Learning — Position Paper Track
A position paper arguing that the dual curse of multilingual AI — 35% harmful generation and near-random reward-model accuracy in low-resource languages — cannot be fixed by post-hoc alignment, and instead requires socio-technical guardrails, pre-training interventions, community-led harm specification, and multilingual safety-usability metrics.
In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), Main Conference
DIA-HARM evaluates 16 harmful content detection models across 50 English dialects using 195K+ samples, revealing 1.4–3.6% F1 drops for fine-tuned models and up to 27% for zero-shot approaches, highlighting equity gaps in content moderation.
In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2026), Datasets and Benchmarks Track
BLUFF is the largest multilingual fake news detection benchmark, spanning 79 languages with 202K+ samples. It introduces AXL-CoI for adversarial generation and mPURIFY for quality filtering, revealing up to 25.3% F1 degradation on low-resource languages.
2025
In Findings of the Association for Computational Linguistics, Empirical Methods in Natural Language Processing
Chain-of-Interactions (CoI) introduces a novel multi-step framework that leverages LLMs' in-context learning capabilities for abstractive task-oriented dialogue summarization. Through comprehensive evaluation across 11 models and human studies with 480 evaluators, CoI demonstrates decisive superiority with 6× better entity preservation and 49% higher quality scores compared to existing approaches, establishing new standards for customer service dialogue summarization.
In Findings of the Association for Computational Linguistics, Empirical Methods in Natural Language Processing
GAMIC introduces a novel self-supervised learning approach for molecular in-context learning that combines graph neural networks with Morgan fingerprints to better capture molecular complexity. By aligning global molecular structures with textual descriptions and using Maximum Marginal Relevance for demonstration selection, GAMIC outperforms traditional Morgan-based retrieval methods by up to 45% across diverse molecular tasks.
In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics, Human Language Technologies
Beemo introduces a novel benchmark featuring 6.5k expert-edited machine-generated texts across diverse domains from creative writing to summarization. Through comprehensive evaluation of 33 MGT detector configurations, we demonstrate that expert editing effectively evades detection while LLM-edited texts remain distinguishable from human writing, highlighting critical gaps in current detection methods for multi-author scenarios.
In Proceedings of the 31st International Conference on Computational Linguistics
This work introduces semantic captioning for SQL queries, addressing the reverse operation of semantic parsing by translating SQL code into natural language explanations. Using graph-aware few-shot in-context learning with smaller LLMs, our approach outperforms random selection by up to 39% on BLEU score, providing crucial capabilities for understanding and explaining LLM-generated SQL code in security-critical applications.
2024
In Findings of the Association for Computational Linguistics, Empirical Methods in Natural Language Processing
This research from Penn State and KiNiT, benchmarks the effectiveness of 10 authorship obfuscation (AO) techniques against 37 machine-generated text (MGT) detection methods across 11 languages, totaling 4,070 evaluations. It reveals that all AO methods can evade detection in every language, particularly highlighting the efficacy of homoglyph attacks. This underscores the need for improved multilingual MGT detection strategies.
2023
In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing
This research project is a collaboration with Penn State and MIT Lincoln Lab. Our study demonstrates the dual capacity of LLMs for offensive misuse and defense detection against disinformation without requiring additional training.
In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing
This research from Penn State and KiNiT introduces MULTITuDE, a novel multilingual dataset for detecting machine-generated text. Comprised of over 74,000 authentic and artificially-generated texts in 11 languages from 8 models, MULTITuDE benchmarks text generation capabilities in non-English languages and multilingual detection performance. The dataset addresses current gaps in analyzing and systematically evaluating machine text generation and detection across multiple languages.
2026
In Proceedings of the 2nd Workshop on Speech and Audio Language Models (SALMA 2026), co-located with EMNLP ‘26
Trilingual spoken hallucination detection, comparing what is detectable from raw audio against what survives transcription.
2022
In Association for Computational Linguistics Proceedings of the Workshop on Combating Online Hostile Posts in Regional Languages during Emergency Situations, 2022 Dublin, Ireland.
This paper is the first attempt to detect COVID-19 misinformation (in English, Spanish, and Haitian French) populated in the Caribbean regions, using the fact-checked claims in the US (in English).