Silent memory for language processing

Fitz, H. 1 , van den Broek, D. 1 , Uhlmann, M. . 1 , Duarte, R. 2, 3, 4, 5 , Hagoort, P. 1, 6 & Petersson, K. M. 1, 6

1 Neurobiology of Language Department, Max Planck Institute for Psycholinguistics Nijmegen, the Netherlands
2 Institute of Neuroscience and Medicine (INM-6), Institute for Advanced Simulation (IAS-6) and JARA BRAIN Institute I, Jülich Research Centre, Germany
3 Bernstein Center Freiburg, Albert-Ludwig University of Freiburg, Germany
4 Faculty of Biology, Albert-Ludwig University of Freiburg, Germany
5 Institute of Adaptive and Neural Computation, School of Informatics, University of Edinburgh, UK
6 Donders Institute for Brain Cognition and Behaviour, Center for Cognitive Neuroimaging, Radboud University Nijmegen, the Netherlands

Integrating meaning over time requires memory ranging from milliseconds (words) to seconds (sentences) and minutes (discourse). How do transient events like action potentials support memory at these different time-scales? Here we investigate the nature of processing memory in a neurobiologically motivated model of sentence comprehension.

The model was a recurrent, sparsely connected network of spiking neurons. As input it received word sequences generated from construction grammar templates and their syntactic alternations (e.g., transfer dative, caused motion). Word durations varied between 50ms and 0.5s of real, physical time. It's task was to incrementally map these sequences onto semantic roles.

To probe memory, we systematically manipulated network connectivity, the shape of synaptic currents, and properties of neuronal adaptation. Near optimal performance was observed when synaptic and neuronal time-constants were tuned to the temporal characteristics of the comprehension task. Recurrent connectivity only played a limited role in maintaining information over time.

These results suggest that memory for language is provided by activity-silent dynamic processes rather than the active replay of prior input as in storage-and retrieval models of working memory. Therefore, the development of neurobiologically realistic, causal models will be critical for our understanding of language processing.