Natural language parsing using simple neural networks

This paper describes a method for performing first order rule-based inference in a massively parallel neural architecture. The emerging system, though not equivalent to full first order logic, is sufficiently rich to implement production rules for parsing natural language. Our system is novel in that, whereas previous network implementations of first order reasoning represented concepts by patterns of activation, our networks represent concepts by specific nodes in the network, allowing greater parallelism of inference. The key to our approach is the use of nodes whose stimulation levels are matrices, rather than scalars.