Part of speech tagging with min‐max modular neural networks

A parts of speech (POS) tagging system using neural networks has been developed by Ma and colleagues. This system can tag unlearned data with a much higher accuracy than that of the Hidden Markov Model (HMM), which is the most popular method of POS tagging. It does so by learning a small Thai corpus on the order of 10,000 words that are ambiguous as to their POSs. However, the three‐layer perceptron used in the system has slow convergence and low learning accuracy even on such a small amount of data. It is therefore difficult to improve accuracy by incrementing the epoch of learning or by increasing the amount of learning data. To solve this problem, the tagging system of this paper makes use of the min‐max modular (M 3 ) neural network of Lu and colleagues. This new system learns faster and has a higher learning accuracy compared with the old one, by decomposing large, complicated POS tagging problems into many smaller, easier problems. Learning accuracy can be improved by using the same learning data and larger data sets can be learned, which results in a much higher tagging accuracy. © 2002 Wiley Periodicals, Inc. Syst Comp Jpn, 33(7): 30–39, 2002; Published online in Wiley InterScience ( www.interscience.wiley.com ). DOI 10.1002/scj.1139