How Can Metaphors Be Interpreted Cross-Linguistically?

Abstract

Research on metaphor as a phenomenon amenable to the techniques of computational linguistics received a substantial boost from a recent US government funding initiative (iARPA, http://www.iarpa.gov/Programs/ia/Metaphor/metaphor.html) that set up a number of teams in major universities to address the issues of metaphor detection and interpretation on a large scale in text. Part of the stated goal of the project was to detect linguistic metaphors (LMs) computationally in texts in four languages and map them all to a single set of conceptual metaphors (CMs). Much of the inspiration for this was the classic work of George Lakoff (Lakoff and Johnson 1980) which posited a set of universal metaphors in use across cultures and languages. I wish to examine the assumptions behind this goal and in particular to address the issue of how and in what representation such CMs can be expressed. I shall argue that a naïve approach to this issue is to make very much the same assumptions as the work of Schank and others in the 1970s: namely that there can be a universal language of “primitives” for the expression of meaning, which in practice always turns out to be a form of simple English. Reviving that assumption for the study of metaphor raises additional issues since, even if the senses of the terms in those CM representations could be added to the representations, metaphors often deploy new senses of words which will not be found in existing sense inventories like computational lexicons. The paper is not intended just to present a negative conclusion; I also argue that the representation of metaphors in a range of languages can be brought together within some CM scheme, but that simply reviving the English-as-interlingua assumptions of 40 years ago is not a good way to make progress in this most difficult area of meaning computation. In what follows I discuss first the representation of CMs: in what language are they stated? I argue the need for some inclusion of the representation of the senses of their constituent terms within the CM, or at least a default assumption that the major sense (with respect to some lexicon such as WordNet) is the intended one. I then consider the issue of conventional metaphor and its representation in established lexicons (again such as WordNet) and the effect that can have on detection strategies for metaphor such as selectional preference breaking. I then argue that the map** of text metaphors to CMs, as well as the empirical, rather than intuitive, construction of CM inventories require further use of preference restrictions in lexicons by means of a much-discussed process of projection or coercion. I conclude that only the use of (computable) procedures such as these for metaphor detection and map** can lead to a plausible program for the large scale analysis of metaphor in text and that Lakoff’s views on metaphor lack these empirical underpinnings.