Modeling sentence production as a staged generative process

Two-stage models of grammatical encoding posit that sentence production unfolds in two sequential steps: the construction of a hierarchical syntactic structure, followed by its linearization into a sequence suitable for articulation. While widely accepted in psycho- and neurolinguistics, such models lack formal computational implementations. Here, we introduce a novel Multinomial Processing Tree (MPT) model that operationalizes this two-stage framework to explain syntactic error patterns in individuals with aphasia. Drawing on discourse samples annotated for distinct error types, we fit an MPT model to estimate individual abilities at each processing stage. These ability estimates correlated with observed error rates and localized to distinct neural substrates: hierarchical encoding ability was linked to the posterior superior temporal sulcus and parietal cortex, linearization to posterior frontal and anterior parietal regions, and omission-related processes to the anterior cingulate cortex. Our findings support a neurocomputational dissociation between hierarchical and linear stages of grammatical encoding, aligning with prior lesion-symptom mapping and theoretical accounts of expressive agrammatism and paragrammatism. This work represents the first computational instantiation of a two-stage model of syntactic production, bridging formal modeling and lesion analysis to advance our understanding of the architecture and breakdown of grammatical encoding.