Neural network for syntactic categorisation of words

This paper demonstrates the results of a neural network based system used to solve the problem of disambiguating words into syntactic categories. This syntactic tagging is very important for both speech synthesis and speech recognition as around 50% of English words can have multiple syntactic functions (or tags) dependent on their context-the syntactic function implies information about the stress and intonation placed upon a word. The approach taken here utilises a fully connected multilayer perceptron (MLP) with the input space split into future, current and previous contexts. Network size and training time was reduced as much as possible to make the potentially huge network more manageable. Overall, the system could distinguish syntactic function to 97% accuracy, with a set of 46 tags, and 93.3%, with a full set of 153 tags from the Lancaster-Oslo/Bergen (LOB) corpus, on unseen data. These results compare very favourably with results obtained by other researchers using statistical and neural network techniques