A new architecture for representing parsing of natural language is described which conforms to psycholinguistic, neurolinguistic and computational constraints. The parsing model uses a particular spreading activation or neural network scheme called connectionism which entails a massive number of appropriately connected computing units that communicate through weighted levels of excitation and inhibition. Such an architecture adds considerable constraints of its own which serve to explain some constraints at the functional level. The model accounts for psycholinguistic data on the access of word meanings, recent neurolinguistic data on agrammatism, and some of the apparent parsing strategies of normals. |