Learning to buy a Renault and talk to BMW: A supervised approach to conventional metonymy

We test the validity of a machine learning approach to metonymy resolution, presenting a case study for organisation names and comparing the results with a previous experiment on location names. We de-scribe a reliable annotation scheme for organisation names and present a corpus annotated for metonymic usage. We then discuss experiments with a supervised classification algorithm on this corpus, focusing on feature contributions and highlighting advantages and disadvantages of a classification approach to metonymy resolution.