How is the generation of a grammatical sentence implemented by the human brain? A starting place for such an inquiry lies in linguistic theory. Unfortunately, linguistic theories illuminate only abstract knowledge representations and do not indicate how these representations interact with cognitive architecture to produce discourse. We examine tightly constrained empirical methods to study how grammar interacts with one part of the cognitive architecture, namely attention. Finally, we show that understanding attention as a neural network can link grammatical choice to underlying brain systems. Overall, our commentary supports a multilevel empirical approach that clarifies and expands the connections between cognitive science and linguistics thus advancing the interdisciplinary agenda outlined by Jackendoff.