A Neural Network Based Dutch Part of Speech Tagger

In this paper a Neural Network is designed for Part-of-Speech Tagging of Dutch text. Our approach uses the Corpus Gesproken Nederlands (CGN) consisting of almost 9 million transcribed words of spoken Dutch, divided into 15 different categories. The outcome of the design is a Neural Network with an input window of size 8 (4 words back and 3 words ahead) and a hidden layer of 370 neurons. The words ahead are coded based on the relative frequency of the tags in the training set for the word. Special attention is paid to unknown words (words not in the training set) for which such a relative frequency cannot be determined. Based on a 10-fold cross validation an approximation of the relative frequency of tags for unknown words is determined. The performance of the Neural Network is 97.35%, 97.88% on known words and 41.67% on unknown words. This is comparable to state of the art performances found in the literature. The special coding of unknown words resulted of an increase of almost 13% for the tagging of unknown words