Description:

When we talk about language in humans and in machines, prediction is a common thread: much evidence indicates that human brains predict during language processing, and prediction is central to language learning in artificial intelligence models. On the other hand, we have the notion of language "understanding", which surely connects to the linguistic notion of meaning, and which stands as a central goal in AI, but which is a challenge to define satisfactorily. While we can safely assume that humans understand language (in addition to predicting it), it is non-trivial to assess whether prediction-based artificial intelligence models possess something that we might consider "understanding". In this talk I will discuss a series of projects analyzing AI models' linguistic representations at the word and sentence levels, as well as the predictive behaviors of such models, drawing comparisons to language comprehension and prediction phenomena in human cognition. I will discuss the implications of these results with respect to what exactly these models learn about language during their training, and how these models' language capacities relate to human processes of meaning understanding versus predictive processing.