Using computational semantics to study meaning in the brain

How do humans understand the meaning of individual words? How do we combine the meaning of multiple words to comprehend novel sentences? Cognitive neuroscience has addressed these questions by studying patterns of neural activation across the cerebral cortex, identifying the regions responsible for language processing. Meanwhile, computational linguistics has approached these questions by constructing vector-based models demonstrating how meaning can be represented as patterns of activation in artificial neural networks. In recent years, increasing recognition of these commonalities, along with the rapidly developing linguistic capabilities of large language models, has led to a surge in research applying computational semantics models to study human language processing in the brain. In this cross-disciplinary review, we aim to better integrate theoretical and experimental approaches by providing an overview of vector-based semantics models, including a summary of major approaches to modelling word and sentence meaning. We next review different approaches to studying semantics in the brain, including traditional approaches focusing on localisation of lexical and compositional semantics, and newer methods using word embeddings and syntactic parsing algorithms. We then discuss provide a comparative analysis of 57 studies utilising vector-based semantics models to study language processing with functional Magnetic Resonance Imaging (fMRI), in which we highlight both common trends as well as significant methodological inconsistencies across studies which hinders interpretation. We conclude by emphasising the value of vector-based semantics models for understanding semantic processing in the brain, while highlighting the need for greater methodological standardisation.