Corpus Based Methods for Learning Models of Metaphor in Modern Greek

In this paper we propose a method for detecting metaphorical usage of content terms based on the hypothesis that metaphors can be detected by being characteristic of a different domain than the one they appear in. We formulate the problem as one of extracting knowledge from text classification models, where the latter have been created using standard text classification techniques without any knowledge of metaphor. We then extract from such models a measure of how characteristic of a domain a term is, providing us with a reliable method of identifying terms that are surprising for the context within which they are used. To empirically evaluate our method, we have compiled a corpus of Greek newspaper articles where the training set is only annotated with the broad thematic categories assigned by the newspapers. We have also manually annotated a test corpus with metaphorical word usage. In our experiment, we report results using tf-idf to identify the literal (characteristic) domain of terms and we also analyse the interaction between tf-idf and other typical word features, such as Part of Speech tags.