Title: Revisiting met*
Speaker: Kevin Mathews (HITS)
The existing metonymy resolution techniques detect only deviation from literal usage. Thus, metonymy resolution is formulated as a binary classification task, where the classes are metonymic and literal. In addition, the principal hypothesis used to detect metonymy is violation of selectional preferences. However, this violation is neither necessary nor sufficient for metonymy to occur as it is also associated with other linguistic phenomena such as metaphor. In this work, the objective is to detect metonymy by establishing metonymic links between the vehicle (metonymic word) and the target (actual interpretation). For this purpose, we investigate a 3-class classifier, where the classes are metonymic, metaphoric and literal. We also describe a method to compute metonymic associations between concepts using the Wikipedia disambiguation pages.