A parallel evolutionary algorithm for stochastic natural language parsing

Abstract

This paper presents a parallel evolutionary program for natural language parsing. The implementation follows an island model, in which, after a number of generations, demes exchange some individuals in a round-robin manner. The population is composed of potential parsings for a sentence, and the fitness function evaluates the appropriateness of the parsing according to a given stochastic grammar. Both the fitness function and the genetic operators, which require that the result of their application still corresponds to the words in the input sentence, are expensive enough to make the evolutionary program appropriate for a coarse grain parallel model and its distributed implementation. The system has been implemented in a parallel machine using the PVM (Parallel Virtual Machine) software. The paper describes the study of the parameters in the parallel evolutionary program, such as the number of individuals to be exchanged between demes, and the number of generations between exchanges. Different parameters of the algorithm, such as population size, and crossover and mutation rates, have also been tested.