Characterizing Language Representations in the Human Mind and Brain

Abstract

Language allows for the mapping of speech signals or written characters to meaning every time we engage in conversation or read. How can biological tissue, our brains, support this mapping process? This thesis characterizes the neural representations that enable humans to infer the meaning of a sentence. My work builds on the foundation that regions in the left frontal and temporal parts of the brain causally and selectively support language processing (the ‘language network’). Chapter 2 asks how the language network develops. Through a case study of an individual born without their left temporal lobe (but with neurotypical language abilities), I demonstrate that the presence of temporal language regions appears to be necessary for the development of ipsilateral frontal regions, which echoes evidence from aphasia that the temporal areas are more important for language function. Chapters 3-5 aim to understand the representations and computations that mediate language comprehension. Traditionally, this line of inquiry has been challenging given the limited utility of probing animal models whose communication systems differ substantially from human language. However, the recent advent of artificial language models (LMs) has demonstrated that a system other than the human brain is capable of generating fluent and coherent text. Chapter 3 introduces the use of LMs as model systems for studying neural representations of language. I ask what aspects of an LM’s representation of the linguistic input matter the most for model-to-brain similarity. Across a series of systematic comparisons, I show that meanings of content words, such as nouns and verbs, matter more than syntactic structure (e.g., word order and function words). In Chapter 4, I leverage this model-to-brain similarity to ask what kinds of linguistic input the human language regions are most responsive to. I use an LM to identify sentences that maximally drive or suppress activity in language regions, and I demonstrate that these regions respond most strongly to sentences that are sufficiently linguistically well-formed but unpredictable in their structure or meaning, suggesting that this network is tuned to input predictability in the service of efficient meaning extraction. Finally, in Chapter 5, I use high-field (7T) fMRI to search for the organizing dimensions of the language network. By performing a data-driven decomposition of neural responses to linguistically diverse sentences, I show that only two components—shared across individuals—emerged robustly, accounting for about 34% of the explainable variance. In line with work in Chapter 4, the first component appears to correspond to processing difficulty. The second component appears to correspond to meaning abstractness. Both components are distributed across frontal and temporal brain areas but show systematic topographies across participants. Altogether, this thesis provides a detailed characterization—across thousands of sentences and through spatially-precise neural measurements—of how the fronto-temporal language network supports language comprehension. This work brings us closer to deciphering the circuits and mechanisms that underlie the astonishing human capacity to infer complex meanings through language.