Recent & Upcoming Talks

The widespread use and disruptive effects of large language models (LLMs) have led to concerns about their potential misuse, such as generating harmful and misleading content on a large scale. To address this risk, the authors propose a novel “Fighting Fire with Fire” (F3) strategy, which utilizes the generative and reasoning capabilities of modern LLMs to counter disinformation created by both humans and LLMs.

In this talk we present research that tackles the misuse of large language models (LLMs) by introducing the Fighting Fire with Fire (F3) strategy, which uses GPT-3.5-turbo to generate and detect disinformation. By employing advanced techniques, we achieved a 68-72% accuracy in identifying deceptive content. We also address COVID-19 misinformation in low-resource regions, focusing on the Caribbean. Using US fact-checked claims, we trained models to detect misinformation in English, Spanish, and Haitian French. Our results highlight the limitations of current methods and the need for further multilingual research.

Michiharu Yamashita and I co-present our latest research, Titled, Fighting Fire With Fire - The Dual Role of Large Language Models in Crafting and Detecting Elusive Disinformation. This is a collaboration with Penn State and MIT Lincoln. Our study demonstrates the dual capacity of LLMs for offensive misuse and defense detection against disinformation without requiring additional training.

This talk explores recent advances in AI and natural language processing. It highlights an influential EMNLP 2023 paper: “Fighting Fire with Fire” - introducing adversarial attack methods that reveal weaknesses in AI systems by inducing model misclassifications. However, it also shows the dual use for combatting harmful text. Published among 900 highly selective papers, this work epitomizes progress in strengthening robustness, while using AI’s dual capacities for good. Brief background will be provided before diving into the state of the field and latest innovations showcased across the cutting-edge EMNLP research.

I co-presented with Cristal Giorio Jackson at the 2023 NSF NRT Annual Meeting in Arizona State University. This NSF-funded research examines whether mismatching native accent variants impacts language processing for native listeners. An EEG experiment found successful semantic access occurred when listeners heard matched native variants, but unsuccessful semantic mapping when hearing unfamiliar, less intelligible variants. This suggests processing costs when speech doesn’t match a native variant. The project integrates psychology, linguistics and AI to explore improving automatic speech recognition and language ID for regional/non-native accents. Techniques like leveraging EEG biomarkers and combining acoustic models with character language models may enhance accent robustness.

This presentation at the “Research in Action” alumni program was delivered alongside Dr. Carol Miller and Suhas Nagaraj and focused on leveraging AI in speech and language therapy for children. It addressed the significant challenges in providing quality speech and language (S&L) services to children with various disorders. The team discussed the potential of AI technologies, like multimodal learning and federated learning, to enhance the availability and effectiveness of these services. They also highlighted the importance of responsible data use, considering privacy and ethical concerns, especially in the context of children’s speech data.

This paper examines the challenge of detecting false claims related to COVID-19 circulating online in Caribbean countries. The Caribbean’s linguistic diversity and scarcity of fact-checking pose difficulties for AI models trained on English data. Experiments found classical machine learning approaches struggle to identify Caribbean false claims. More sophisticated techniques like transfer learning with multilingual models showed promise. But all models suffered reduced performance on non-English content. The paper concludes AI solutions developed for high-resource settings have limited portability to low-resource regions facing endemic barriers around data, language, and misinformation ecosystems.

I present my latest research, Titled, Detecting False Claims in Low-Resource Regions - A Case Study of Caribbean Islands. Our study democratize the challenges in transferling and applying Anglo-centric models in low-resource settings, evening with human writen or machine translated english training data.

My research presentaton addresses the challenge of detecting false claims in low-resource regions, with a focus on the Caribbean Islands. The presentation highlighted the limitations of high-resource language detectors in such regions, the inadequacy of classical machine learning models, and the potential of state-of-the-art models with limited generalizability. Lucas emphasized the importance of pretrained language models and transfer learning in enhancing downstream detectors, advocating for multilingual fake news detection and cross-lingual generalizable approaches.