What compositionality reveals about meaning, cognition, and machines

This perspective article discusses the nature of linguistic computation in cognition vis-à-vis linguistic meaning. In particular, this article explores the computational basis of linguistic meaning from a compositional computability perspective and then shows that linguistic meaning is not computable under certain conditions of compositionality, which requires the dependence of the meaning of a whole expression on the meanings of the parts. In other words, compositionality does not guarantee the computability of linguistic meanings. This is also partly governed by the syntactic underdetermination of linguistic meanings as conceived of within syntactic theories assuming computability of structural operations for compositional mapping. So, syntax, understood in a purely computational sense, does not guarantee computability of semantic composition when understood in the context of a cognitive mechanism of understanding linguistic meaning. The limits of computability of linguistic meaning are then indicated. This is investigated from the vantage point of unification of cognitive and formal aspects of linguistic meaning in brain dynamics. This turns out to have serious consequences for Large Language Models (LLMs)/Generative AI that appear(s) to stumble over linguistically deep or pragmatically rich aspects of meaning. Thus, this article projects insights into the limits of computational basis of linguistic meaning in humans and machines.