Markov Chain Monte Carlo Methods and Evolutionary Algorithms for Automatic Feature Selection from Legal Documents

摘要

In this paper, we present three different approaches for feature selection, starting from a na?ve Markov Chain Monte Carlo random walk algorithm to more refined methods like simulated annealing and genetic algorithms. It is typical for textual data to have thousands of dimensions in their feature space which makes feature selection a crucial phase before the final classification. Classification of legal documents into eight categories was performed via a simple document similarity measure based on term frequency and the nearest neighbour concept. With an average success rate of 76.4%, the random walk algorithm not only performed better than the simulated annealing and genetic algorithms but also matched the accuracy of support vector machines. Although these methods have commonly been used for selecting appropriate features in other fields, their use in text categorisation have not been satisfactorily investigated. And, to our knowledge, this is the first work which investigates their use in the legal domain. This generic text classification framework can further be enhanced by using an active learning methodology for the selection of training samples rather than following a passive learning approach.