Connectionist unification of feature-structures

Summary form only given, as follows. A general method is presented to encode and unify recursively nested feature structures in neural nets. The unification algorithm implemented by the net relies on iterative coarsening of equivalence classes of graph nodes. This approach allows the reformulation of unification as a constraint satisfaction problem and allows the connectionist implementation to take full advantage of the potential parallelism inherent in unification. Moreover, the method is able to process any number of feature structures in parallel, making decisions among mutually exclusive unifications where necessary.<>