Conference on Cognitive Computational Neuroscience

Our memory for sentences is influenced by long-term knowledge of syntactic structures. However, no computational model of working memory can yet account for syntax. Here we propose a neural architecture involving two types of units: a set of word units, and a set of conjunctive units for abstracting syntactic information. The synapses undergo Hebbian plasticity to allow rapid changes in connections between the word and conjunctive units, as well as the connections between conjunctive and conjunctive units. We will show that our model is able to 1) remember a sequence of words better when they form a syntactic sentence compared to a shuffled list, 2) accommodate words with multiple syntactic roles, and 3) switch between branching structures of syntactically different sentences. These abilities arise from the rapid creation of new attractor states, and permit manipulation of information within working memory. This is the first computational model to embed syntax into working memory, and provides a neural basis for sequential symbolic cognition.