Extraction of Nominal Multiword Expressions in French
Marie Dubremetz, Joakim Nivre · 2014
Multiword expressions (MWEs) can be extracted automatically from large corpora using association measures, and tools like mwetoolkit allow researchers to generate training data for MWE extraction given a tagged corpus and a lexicon. We use mwe-toolkit on a sample of the French Europarl corpus together with the French lexicon Dela, and use Weka to train classifiers for MWE extraction on the generated training data. A manual evaluation shows that the classifiers achieve 60–75 % precision and that about half of the MWEs found are novel and not listed in the lexicon. We also investigate the impact of the patterns used to generate the training data and find that this can affect the trade-off between preci-sion and novelty. 1