Should LLM technology change the way we think about language itself? This project conducts a philosophical investigation into whether LLMs reveal fundamental insights about language. While some argue that LLMs can function as descriptive and predictive accounts of language comparable to those developed by linguists, others have claimed that they provide no insight into human language whatsoever. This project will develop a new middle ground: that LLMs can serve as scientific models of public languages.
The research is interdisciplinary in nature, combining philosophical analysis informed by recent debates in philosophy of language and philosophy of science with empirical investigations of LLM technology that makes use of recent innovations in mechanistic interpretability and explainable AI. As such, the project will establish links between philosophy, computational linguistics, and AI research that have until now been underexplored. The potential for engagement with relevant stakeholders in industry is also an important part of the project. Understanding how LLMs can be sources of understanding with regard to language has the potential to impact questions of trust and reliability that are often paramount to those who develop or employ such systems.
The project is a continuation and expansion of Grindrod's recent research, notably his 'Modelling Language using Large Language Models', 'Transformers, Contextualism, and Polysemy', 'Large Language Models and Linguistic Intentionality', 'Distributional Theories of Meaning', and 'Distributional Semantics, Holism, and the Instability of Meaning' (co-authored with J.D. Porter and Nat Hansen).
The postholder will collaborate with the PI on model analysis studies investigating how specific linguistic phenomena are processed in large language models. Essential expertise in areas such as: NLP, mechanistic interpretability, or computational linguistics.