All are invited to this Philosophy colloquium talk at 1:30 PM on Wednesday 27 August in AC04 LR302.

Title: Scope Ambiguities in Large Language Models

Speaker: Gaurav Kamath

Abstract: Large language models (LLMs) have had an immense impact on society at large through applications like ChatGPT. But what do we know about how they process formal structures in language, and how this compares with humans? Conversely, what can these models tell us about human language? In this talk, I will focus on two research projects that attempt to tease at both questions. The first project looks at how LLMs process scope ambiguities — interpretative ambiguities in sentences, typically containing quantifiers, that can receive multiple translations to formal logic. The second project uses language models to answer questions about how words change meaning over time, and individuals' capacities to adopt these changes in adulthood. In both cases, we find strong connections between LLMs and the study of human language: LLMs show strong correlations with humans in their interpretations of scope ambiguity, while applying them to study word meaning change reveals novel insights about human linguistic behavior.  

Bio: Gaurav Kamath is a 4th-year Linguistics PhD student at McGill University and Mila — Quebec AI Institute, under the supervision of Professor Siva Reddy. His research focuses on assessing the linguistic capacities of large language models, and using them to answer questions about human language. His work has been published in leading journals, including Transactions of the Association for Computational Linguistics and Proceedings of the National Academy of Sciences, and presented at international linguistics and AI conferences. Gaurav was a member of the UG'19/ASP'20 batch at Ashoka University, majoring in Politics, Philosophy and Economics.