A Neural Model of Compositional Sentence Structures

A neural architecture for compositional sentence structures is presented. The architecture solves the ‘four challenges for cognitive neuroscience’ described by Jackendoff (2002). Sentence structures are encoded in this neural architecture by temporarily binding word representations with structure representations in a manner that preserves sentence structure. The architecture can store different sentence structures simultaneously. Answers to specific ‘who does what to whom’ questions can be produced by means of a selective activation process within the architecture. The architecture can account for effects of sentence complexity. Introduction Jackendoff (2002) analyzed the difficulties given by the neural instantiation of linguistic structures (‘four challenges for cognitive neuroscience’), in particular the binding problem in language and the problem of multiple instantiation (the ‘problem of 2’). Figure 1 provides an illustration. A word like cat will activate the same neural structure (or ‘word assembly’) in The cat chases the mouse and The mouse chases the cat. This raises the question of how the multiple instantiations of cat (and mouse) are distinguished in both sentences (the problem of 2) and how cat and mouse are bound correctly to chases, without creating the incorrect bindings of cat chases cat and mouse chases mouse (the binding problem). Figure 2 illustrates a solution of these problems, in which neural ‘word assemblies’ are embedded in a neural architecture consisting of neural ‘structure assemblies’ and neural ‘memory circuits’ used for binding. A structure assembly consists of a ‘main assembly’ and a number of ‘subassemblies’. Two kinds of structure assemblies are shown: verb-phrases (VPs) and noun-phrases (NPs), with agent (a) and theme (t) subassemblies. The subassemblies are connected to the main assembly by gating circuits, which can be activated when certain structural control conditions are met. During syntactic processing, word and structure assemblies are bound to one another by activating memory circuits that connect the assemblies. Figure 3 illustrates a gating circuit, consisting of a ‘disinhibition’ circuit that controls the activation flow between two assemblies (X,Y). X activates an inhibitory neuron ix, which blocks Y’s activation. When an external control circuit activates Ix to inhibit ix, X can activate Y. A similar circuit controls Y’s activation of X. Memory circuits temporarily bind assemblies. Figure 3 shows how X binds to Y (a similar circuit binds Y to X). Memory circuits are gating circuits with control given by a ‘delay’ assembly, which is activated when X and Y are simultaneously active (see Van der Velde 2003). As long as the delay assembly is active, due to self-sustained (reverberating) activation, activation can flow between X and Y, which temporarily binds (merges) these assemblies. N1