A computational, corpus-based metaphor extraction system

This thesis describes a corpus-based, computational system for detecting and analyzing conventional metaphors. The system accomplishes this by finding large-scale systematic variation in selectional preferences across domains.; Selectional preferences are inferred from large, domain-specific corpora. Since no dedicated electronic corpus is big enough to encompass every or even most topics, the system obtains its corpora by dynamically mining the Internet. Selectional preferences are represented by clusters of nodes from the WordNet ontology [Miller 90].; The essence of metaphor is transferring structure from a source domain to a target domain. A necessary component of structure transference is metaphoric equivalence between two disparate sets of objects in the source and target domain. This transference has an consequence that can be detected by a purely syntactic analysis. Namely, that the predicates selecting for an object in the source domain tend to select for its equivalent in the target domain, even if those predicates are otherwise atypical of the target domain.; The competence of other computational metaphor systems is generally limited to a handful of instances by dependency on small, hand-coded semantic knowledge bases. The system described in this thesis has no semantic knowledge (beyond WordNet), yet it can detect most conventional metaphors that have concrete source and target domains. The system can not detect rare and novel metaphors.; The system is tested on its ability to derive a subset of the Master Metaphor List [Lakoff, Espenson & Schwartz 91], a manually assembled catalog of conventional metaphors.