Application of genetic programming to text categorization.

摘要

This dissertation uses genetic programming in text categorization problems. Genetic programming algorithms are applied to a set of news articles to evolve programs that determine whether the article belongs to a particular category. The programs are randomly generated from the set of initial functions and constants. Programs with the fewest amount of false assignments are favored in the selection for recombination in the subsequent iterations of the genetic programming algorithm.; The form of the solution is not determined a priori as in other text categorization methods. The basis set of functions and constants used by the genetic analysis program are specified in advance and may include the three basic logical functions and a set of vocabulary words. Other sets of basis functions can be supplied to the genetic algorithm to obtain different programs. The form in which these functions and constants are combined is determined randomly by the genetic algorithm.; The results indicate that genetic programming methods are in the cases examined as good and slightly better than other decision tree or rule induction methods described by Apté et. al. [Apté 1994]. The Genetic Programming methods used a simpler set of features and functions: no word stemming no explicit stop word removal, local dictionary, Boolean functions. The F ₁-measure of categorization performance of 80.4% achieved by Genetic Programming compares favorably with 78.5% Breakeven performance of traditional Boolean rule induction methods. It is comparable with 80.5% Breakeven performance of the rule induction methods with a more complex feature set such as word frequency [Apté 1994].; Characteristics of Genetic Programming text categorization were studied to understand the sensitivity of Genetic Programming methods to vocabulary size, population size, training and testing set selection methods. Temporal characteristics of the Reuters Article Corpus [Lewis-21578) were studied. The results are of interest to both Genetic Programming as well as Traditional categorization methods and may point to significant future performance improvements in both domains. In some cases these results were better than Apté's.