Answers are grounded in this paper and cite their sources.
A major current approach in theoretical linguistics proposes that there is only a partial, defeasible correspondence between syntax and semantics: meaning and grammar are autonomous, parallel components in the architecture of language. This fact still lacks a satisfactory account. One type of explanation could be provided in terms of computational capability: architectures that separate out meaning and grammar into two processing streams might perform better in tasks that require the system to attend to syntactic vs semantic aspects of sentences. We compare single-stream vs dual-stream transformer architectures, which we train and test on data sets from two event-related brain potential (ERP) studies, with a variety of syntactically and semantically challenging sentences. The single-stream model showed greater sensitivity (the probability that a given sentence s from set S is labelled as belonging to S) with syntactic anomalies, whereas the dual-stream model showed greater sensitivity with semantic anomalies. However, overall, the two models classified sentences with similar accuracy: there were no general advantages for either architecture. Our results are a counterexample to the notion that the functional separation of meaning and grammar in the architecture of human language can be explained in strictly computational terms.
No takes yet. Share an insight, caveat, or question.
Michalon et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: