Part of speech tagging with min‐max modular neural networks

Abstract

Part of speech tagging systems using neural networks have been proposed by Ma, et al. They can tag the untrained data at a practical level of accuracy by training a small Thai corpus with ten thousand order words. The multilayer perceptron (MLP) type of neural networks used, however, was found to converge slowly and took a very long time to train even the above mentioned small amount of training data. This paper presents an alternative method for solving the POS tagging problems with the min-max modular neural network proposed by Lu and Ito. By using this modular neural network, the part of speech tagging problems can be broken down into a number of independent smaller and simpler subproblems, and all of the subproblems can be learned by small network modules in parallel.