Corpus-driven Metaphor Harvesting

Proceedings of the Workshop on Computational Approaches to Figurative Language, pages 5–12, Rochester, NY, April 26, 2007. c⃝2007 Association for Computational Linguistics Corpus-driven Metaphor Harvesting Astrid Reining Institute of Romance Languages University of Hamburg 20146 Hamburg, Germany astrid.reining@uni-hamburg.de Birte L¨onneker-Rodman International Computer Science Institute 1947 Center Street, Suite 600 Berkeley, CA 94704, USA loenneke@icsi.berkeley.edu Abstract The paper presents a corpus-based method for finding metaphorically used lex- emes and prevailing semantico-conceptu- al source domains, given a target domain corpus. It is exemplified by a case study on the target domain of European politics, based on a French 800,000 token corpus. 1 Introduction This investigation is situated within the frame- work of the Hamburg Metaphor Database1 (HMD) (L¨onneker and Eilts, 2004), which collects manual annotations of metaphors in context. HMD anno- tation terminology refers to cognitive linguistic ac- counts of metaphor. These suggest that abstract “target” concepts are often thought and talked of in terms of less abstract “source” concepts (Section 2). On these accounts, the paper presents a method for finding metaphorically used lexical items and char- acterizing the conceptual source domains they be- long to, given a target domain corpus. After mentioning related work on metaphor an- notation (Section 3), we exemplify our method by a case study on the target domain of European pol- itics, for which a French 800,000 token corpus is prepared and imported into a corpus manager (Sec- tion 4). Using corpus manager functions, a small set of highly salient collocates of Europe are classi- fied as candidates of metaphorical usages; after as- sessing their metaphoricity in context, these lexemes 1http://www1.uni-hamburg.de/metaphern are grouped into semantico-conceptual domains for which, in a final step, additional lexical instanti- ations are searched (Section 5). Two important source domains (BUILDING and MOTION) are de- tected, which are supported by over 1,000 manual corpus annotations. The domains can be charac- terized as small networks of EuroWordNet synsets (nodes) and lexical as well as conceptual relations (Section 6). Section 7 concludes the paper. 2 Theoretical Aspects The Conceptual Theory of Metaphor (CTM) worked out originally by (Lakoff and Johnson, 1980) claims that conceptual metaphors such as GOOD IS UP and TIME IS MONEY structure the way we think and influence the way we use language. Concep- tual metaphors are mappings between conceptual domains, for example between the target domain GOOD and the less abstract source domain UP, or between TIME (target) and MONEY (source). Conceptual metaphors are rarely directly referred to in speech or writing: Whereas time is money is a standing expression in English, this is much less so for many other conceptual mappings (cf. ?good is up). Consequently, corpus analysis cannot have as a goal finding conceptual mappings as such. Rather, it can find their manifestations through non-literal us- ages of lexical items – i.e., contexts in which source domain words are used to refer to elements in the target domain. For example, high (a word from the UP source do- main) means ‘good’ in the expression high marks; and spend …

Content truncated to 500 words