Strategies of unsupervised learning for a parallel parsing architecture

A study concerning the analysis of natural language is reported, involving a serial and parallel processing principle already applied to pattern recognition tasks: guided propagation. The network of processing units which supports this principle is connected in the course of processing, due to basic differentiation and generalization mechanisms. It is shown how these basic learning mechanisms can be adapted so as to serve two purposes: the automatic clustering of unexpected words and the extraction of lexical substructures from examples of sentences through different possible strategies. Compared with classical natural language parsers the network that results front the learning mechanisms presented here implements a deterministic, left-to-right, breadth-first parser, the system follows every path simultaneously. in synchrony with the input flow. It works bottom-up from the input data. with a top-down trend, as pathways are reinforced. At a given time, part of the network tends to work in a top-down mode, while the other part works bottom-up, depending selectively on the `familiarity´ of the stored items