An algebraic approach to learning in syntactic neural networks

Abstract

The algebraic learning paradigm is described in relation to syntactic neural networks. In algebraic learning, each free parameter of the net is given a unique variable name, and the net output is then expressed as a sum of products of these variables, for each training sentence. The expressions are equated to true if the sentence is a positive sample and false if the sentence is a negative sample. A constraint satisfaction procedure is then used to find an assignment to the variables such that all the equations are satisfied. Such an assignment must yield a network that parses all the positive samples and none of the negative samples, and hence a correct grammar. Unfortunately, the algorithm grows exponentially in time and space with respect to string length. A number of ways of countering this growth, using the inference of a tiny subset of context-free English as a example, are explored.<>