Constituent structure and the binding problem

Colin Phillips, Matthew Wagers
2006 Behavioral and Brain Sciences  
Van der Velde & de Kamps' model encodes complex word-to-word relations in sentences, but does not encode the hierarchical constituent structure of sentences, a fundamental property of most accounts of sentence structure. We summarize what is at stake, and suggest two ways of incorporating constituency into the model. We are impressed by van der Velde & de Kamps' attempt to take seriously the challenge of capturing the complexity of human language in a neurally plausible model. Their model makes
more » ... . Their model makes it possible to ask questions about the encoding of the details of sentence structure that it was difficult to even ask previously. This is no mean achievement. Nevertheless, we are concerned that the authors' model avoids one of the most fundamental properties of sentence structure, and that this could seriously restrict the scope of the model. Although many of the figures in the target article bear a superficial resemblance to the phrase structure trees of linguistics, the sentence structure representations in the neural model lack the hierarchical constituent structure encoded in phrase structure trees. Phrase structure trees encode bindings between primitive elements (words) that create constituents and also bindings between constituents that form larger constituents. In van der Velde & de Kamps' model, in contrast, only bindings between the basic word-level structural assemblies are encoded. A verb's theme sub-assembly may be temporarily bound to a noun's theme sub-assembly to form the equivalent of a simple verb phrase, but the verb phrase does not itself combine with other sub-assemblies to form larger constituents. The 'S' and 'C' structural assemblies that are employed in the encoding of main clauses and embedded clauses respectively do not delimit clause-sized constituents. Rather, they are word-level structural assemblies whose sub-assemblies bind with the sub-assemblies of other word-level units. The binding of words and phrases to form hierarchically organized constituent structures is a property shared by a wide variety of linguistic models that differ in many other respects (e.g.
doi:10.1017/s0140525x06349023 fatcat:knxmidll4ncjrdle3rupnydlqe