Embodied theories of language posit that the human brain’s adaptations for language exploit pre-existing perceptual and motor mechanisms for interacting with the world. In this paper we propose an embodied account of the linguistic distinction between singular and plural, encoded in the system of grammatical number in many of the world’s languages. We introduce a neural network model of visual object classification and spatial attention, informed by a collection of findings in psychology and neuroscience. The classification component of the model computes the type associated with a visual stimulus without identifying the number of objects present. The distinction between singular and plural is made by a separate mechanism in the attentional system, which directs the classifier towards the local or global features of the stimulus. The classifier can directly deliver the semantics of uninflected concrete noun stems, while the attentional mechanism can directly deliver the semantics of singular and plural number features.

Event-related potentials were recorded to investigate the mechanisms of hierarchical processing and level-repetition effect. Participants identified targets that appeared at global, local, or both levels of hierarchical patterns. Reaction times showed global precedence and level-repetition effects. An occipital P1 wave was enhanced to local relative to global targets. The P1 to local targets was also larger when preceded by global than local targets. Global and both-level target selections were indexed by two posterior negativities peaking at 130 and 190 ms poststimulus, whereas local target selection was indexed by a broad occipitotemporal negativity. A late selection positivity was observed over the left occipitotemporal site for global targets but over the central site for local targets. The findings suggest that sensory-perceptual mechanisms contribute to global precedence and level-repetition effects in hierarchical processing.