News & Events

Meliora Weekend Language Science Lecture

Understanding natural language underlies many of the tasks we might expect a large language model (LLM) to help with—whether it's finding videos to better understand a specific topic, assessing whether a claim is true or not, or using computational tools to solve a particular problem. However, these are still quite challenging, as they require the abilities to handle ambiguities and hidden assumptions in the input or problem statement; interpret complex logical and temporal concepts; and reason across modalities and levels of abstraction. Improvements in basic LLM capabilities lead to opportunities to better address these challenges. Specifically, I will show how we can identify implicit knowledge relevant to the given understanding task-at-hand, and then how we can develop techniques to appropriately expand and "fill in" this knowledge, making it explicit. This talk will show the effectiveness and generalizability of this idea across modalities—such as by eliciting temporal event knowledge to improve zero-shot multi-lingual text-to-video retrieval; and across domains—such as by iteratively targeting and decomposing logical assumptions to improve  "general knowledge" and scientific-based claim verification and tool-based reasoning.

Frank Ferraro is an associate professor in computer science at the University of Maryland, Baltimore County (UMBC). He develops models and systems that capture common intuitions and expectations about what happens (or doesn’t) in complex situations. He does this through novel machine learning approaches, development of computational event semantics methods and resources, and multi-modal learning approaches. Prior to UMBC, he received his PhD from Johns Hopkins University and his Bachelors from the University of Rochester. Learn more about Frank Ferraro.