Neuronal memory for language processing

Abstract

In language processing, an interpretation is computed incrementally within memory while utterances unfold in time. Here, we investigate the nature of this processing memory in a spiking network model of sentence comprehension. We show that the history dependence of neuronal responses endows circuits of biological neurons with adequate memory to assign semantic roles and resolve binding relations between words in a stream of language input. A neurobiological read-write memory is proposed where short-lived spiking activity encodes information into coupled dynamic variables that move at slower timescales. This state-dependent network does not rely on persistent activity, excitatory feedback, or synaptic plasticity for storage. Instead, information is maintained in adaptive neuronal conductances and can be accessed directly during comprehension without cued retrieval of previous input words. This work provides a step towards a computational neurobiology of language.