Neural Networks for Learning Grammars

The straightforward mapping of a grammar onto a connectionist architecture is to make each grammar symbol correspond to a node and each rule correspond to a pattern of connections. The grammar then expresses the competence' of the network. The (unsupervised) grammatical inference problem is therefore: how can a network learn to configure itself to reflect the syntactic structure in its input patterns? I adopt the following learning principles. 1. The network learns by generating its own patterns (dreaming') and comparing their statistical properties with real input (waking') patterns. 2. Discrepancies between dreaming and waking prompt the growth of new nodes. 3. The role of each node is to reduce one aspect of the difference between dreaming and waking. 4. The network periodically tests the contribution of each node and removes the redundant ones. What makes this distinctively connectionist is that all tasks (parsing and generating patterns, sprouting and removing nodes) must be distr..