1999 Volume 14 Issue 6 Pages 1116-1124
A multi-neuro tagger that uses variable lengths of contexts and weighted inputs with information gains for part of speech tagging has been proposed by the authors [Ma 98, 馬99]. It has been shown that the tagger has an accuracy higher than any of those obtained using the single neural networks with the fixed length of inputs, which indicates that the length of the context need not be chosen empirically; it can be selected dynamically instead. In this work, by introducing elastic inputs the tagger is slimed down into a single neural network which inherits the features that the multi-neuro tagger has. Computer experiments show that the new neuro tagger has an accuracy slightly higher than the original one instead. A series of comparative experiments for the new neuro tagger and various probability models, which include the frequency model (a base-line model), local n-gram model, and HMM, are further performed for evaluating the neuro tagger. These experiments show that the elastic neuro tagger is definitely far superior to these probability models.