In a production system, the match subsystem is responsible for triggering production rules in accordance with the data elements in the working memory. Traditionally, it performs an exhaustive search or employs a Rete network in the production memory to recognize triggered production rules whenever a data element appears. This process can be dramatically improved if it involves parallel, associative retrieval techniques for matching. Neural networks, with their renowned capabilities in learining, parallel processing, and associative mapping, seem to be an inherent candidate for solving the match problem. We propose two neural network architectures involving feedforward and probabilistic neural network models to simulate the match subsystem. These architectures learn and record the triggering relationships between data elements and production rules in the network links. They not only use less memory than the Rete network, but also execute faster than the Rete algorithm.