Beyond Tag Trigrams: New Local Features for Tagging

Proceedings of the Third International Conference on Language Resources and Evaluation (LREC 2002)

Abstract

The set of features used by any predictive model is of pivotal importance to its performance. In this paper we show the utility and quantify the effect of adding features consisting of arrangements of words and tags (selected by an expert grammarian) in the local context of a trigram tagger. We look in detail at the effect, on tagging with a large syntactic and semantic tagset, of adding these features. We show that the addition of a set of such features improves the the error rate of a trigram tagger by approximately 11%.