Metonymy and Metaphor: Boundary Cases and the Role of a Generative Lexicon

Recent advances in language technologies in areas such as information extraction (MUC1), question answering (TREC2), and summarization (DUC3) have explored extensively the power of statistical NLP in the age of ready availability of large, annotated datasets. Groundbreaking efforts of organizers of these and other continued shared tasks in the spirit of shared resources drive to make developed tools robust and available to the community. We can rely thus on annotated corpora (LDC),4 taggers (Brill 1995; Hepple 2000) and parsers (Grinberg et al. 1995; Briscoe et al. 2006), and development environments such as GATE (Cunningham 2002) which comes with extensive word lists for named entity recognition and partial analysis tools, such as date grammars, etc. The overall goal is not to understand and represent the meaning of texts in an application-neutral way, but to “get at” bits of information hidden in textual documents and extract as many of them as possible for a particular application, leading to partial analysis methods. While partial “syntactic” analysis (phrase chunking, tagging, etc.) has been a mainstay, partial semantic techniques are still in their infancy, with the notable exception of word sense disambiguation (Edmonds and Kilgarriff 2003), annotation and normalization of temporal information (Verhagen et al. 2005), the detection of