Learning inflectional paradigms using parallel corpora

Abstract

This thesis examines the learning of French inflectional paradigms using parallel corpora. An unsupervised method of accomplishing this is presented, and is contrasted with two other methods which make use only of monolingual corpora. It is found that the combination of all three methods yields the best overall results. The combined model achieves final scores of 54.1% precision, 70.3% recall, and 61.1% f-measure when evaluation against a manually-created gold standard is performed. The use of parallel corpora is found to introduce a performance gain with respect to the learning of irregular morphology.