(New) 11 February 2025, Computational Linguistics Seminar, Martha Lewis

Recent neural approaches to modelling language and concepts have proven quite effective, with a proliferation of large models trained on correspondingly massive datasets. However, these models still fail on some tasks that humans, and symbolic approaches, can easily solve. Large neural models are also, to a certain extent, black boxes - particularly those that are proprietary. There is therefore a need to integrate compositional and neural approaches, firstly to potentially improve the performance of large neural models, and secondly to analyze and explain the representations that these systems are using. In this talk I will present results showing that large neural models can fail at tasks that humans are able to do, and discuss alternative, theory-based approaches that have the potential to perform more strongly. I will give applications in language, reasoning, and vision. Finally, I will present some future directions in understanding the types of reasoning or symbol manipulation that large neural models may be performing.