Quantitative Approaches to Metonymy

Introduction Recent years have witnessed an upsurge of interest in metonymy. From cognitive to computational linguistics, researchers have finally realized that metonymy is ubiquitous in everyday language and that it constitutes an important focus of research. In cognitive linguistics, this has given rise to detailed studies of metonymy as a cognitive phenomenon (Kövecses and Radden, 1998; Peirsman and Geeraerts, forthc), while in computational linguistics, it has sparked interest in how computers could learn how to recognize and interpret metonymical language (Markert and Nissim, 2002; Nissim and Markert, 2003). In both fields, the study of metonymy displays clear gaps, however. First, while cognitive linguistics sees itself as a usage-based discipline, a detailed corpus-based study of metonymy has not yet been conducted. Second, the wide application of the metonymy recognition algorithms in computational linguistics is compromised by their complexity and their reliance on a large number of annotated training examples. The quantitative study of metonymy that is presented here draws on well-known techniques from the field of machine learning in order to address these two problems. First, it uses decision trees in order to determine what context variables determine the reading of a possibly metonymical word. Second, it combines ‘lazy’ learning with active learning in order to reduce complexity and annotation effort in automatic metonymy recognition.