A Parallel Distributed Approach to Parsing Natural Language Deterministically
Stan C. Kwasny · Open Scholarship Institutional Repository (Washington University in St. Louis) · 1988
The Determinism Hypothesis (Marcus, 1980) has given rise to much debate. The hypothesis makes explicit the idea that Natural Language interpretation need not depend in any fundamental way on the use of pseudo-parallelism or backtracking. We are exploring the consequences of this hypothesis in attempting to develop approaches to parsing which integrates current work in parallel distributed adaptive networks. We follow the basic approach of "Wait-and-See" parsing (WASP) which has shown the Natural Language interpretation of all but some varieties of "garden-path" sentences can be deterministically performed using a stack, a buffer for sentence constituents, and partitioned packets of rules. Specifically, we replace the rule packets with a single neural network and train the network with an appropriate training set. Training sets derive from either examples of existing WASP grammars or from traces of sentence processing.