In this paper we present an application of the stacking technique to a chunking task: named entity recognition. Stacking consists in applying machine learning techniques for combining the results of different models. Instead of using several corpus or several tagger generators to obtain the models needed in stacking, we have applied three transformations to a single training corpus and then we have used the four versions of the corpus to train a single tagger generator. Taking as baseline the results obtained with the original corpus (F_(β=1) value of 81.84), our experiments show that the three transformations improve this baseline (the best one reaches 84.51), and that applying stacking also improves this baseline reaching an F_(β=1) measure of 88.43.
展开▼
