Metaphor Interpretation as Selective Inferencing

Metaphor pervades na tu ra l language discourse. This paper describes a computational approach to the i n te rp re ta t i on of metaphors. It is based on a na tu ra l language processing system that uses the discourse problems posed by a tex t to se lect the re levant inferences. The problem of i n t e r p r e t i n g metaphors can then be t rans la ted into the problem of se lec t i ng the re levant inferences to draw from the metaphorical expression. Thus, a metaphor is f requent ly given a cor rect i n te rp re ta t i on as a byproduct of the other things a natura l language system has to do. Two examples of metaphors are given — a s p a t i a l metaphor schema from computer sc ience, and a novel metaphor — and it is shown how the i n t e r p r e t a t i o n problem fo r each can be t rans la ted i n t o a se lec t i ve ln ferenc ing problem and solved by the ord inary operat ions of the system. This framework sheds l i g h t on the analog ica l processes that under l ie metaphor and begins to exp la in the power of metaphor. 1. Metaphor is Pervasive I . A. Richards, in speaking of metaphor, s a i d , " L i t e r a l language is rare outside the cen t ra l parts of the sc iences. " (Richards 1936). But i t is rare even in the cen t ra l parts of the sciences. Consider f o r example the fo l l ow ing tex t from computer sc ience. It comes from an a lgor t ihm desc r ip t i on in the f i r s t volume of Knuth's Ar t of Computer Programming, Vo l . 1, p. 417, and is at but one remove from the domain's most formal mode of expression. "Given a po in ter PO, t h i s a lgor i thm sets the MARK f i e l d to 1 in NODE(PO) and in every other node which can be reached from NODE(PO) by a chain of AL1NK and BLINK po in ters in nodes w i t h ATOM MARK = 0. The a lgor i thm uses three pointer va r i ab les , T, Q, and P, and modif ies the l i n k s and con t ro l b i t s dur ing i ts execut ion in such a way tha t a l l ATOM, ALINK, and BLINK f i e l d s are restored to t h e i r o r i g i n a l se t t i ngs a f t e r complet ion, al though they may be changed temporar i l y . " In t h i s tex t the a lgo r i t hm, or the processor that executes i t , is apparent ly a purposive agent tha t can perform such act ions as rece iv ing po in te r s ; s e t t i n g , changing, and res to r i ng f i e l d s ; reaching nodes; using var iab les fo r some purpose; modify ing l i n k s and b i t s ; and executing and completing i t s task . Nodes are apparent ly locat ions that can be l i nked and st rung i n t o paths by po in ters and v i s i t e d by the processor-agent. Nodes a lso seem to be containers that can conta in f i e l d s . F ie lds are a lso containers tha t can conta in p o i n t e r s , among other t h i ngs . In a d d i t i o n , f i e l d s are e n t i t l e s tha t can be placed a t , or set t o , locat ions on the number scale or in the s t ruc tured t h e i r very name, suggest ob jects that can po in t to a l oca t i on f o r some agent 's i n fo rmat ion . In f a c t , there i s very l i t t l e in the paragraph that does not res t on some s p a t i a l or agent metaphor. Moreover, these are not simple i so la ted metaphors; they are examples of la rge-sca le "metaphor schemata" or " roo t metaphors" (Lakoff and Johnson 1980) which we use to encode and organize our knowledge about the objects of computer sc ience. They are so deeply ingra ined tha t t h e i r metaphorical character genera l ly escapes our no t i ce . The pervasiveness of metaphor was noted as ear ly as the e ighteenth century by Giambatt is ta Vico (1744, 1968). In our century , t h i s observat ion has been the basis f o r a r e j e c t i o n of A r i s t o t l e ' s and Q u i n t i l l i a n ' s views tha t metaphor is mere ornament, and an e leva t ion of metaphor to an "omnipresent p r i n c i p l e of language" (Richards 1936) and " the law of i t s l i f e " (Langer 1942). As we saw in our example, the s p a t i a l metaphor espec ia l l y is pervas ive. Jespersen (1922) remarked on t h i s . For Whorf (1939) it was a key element in h is view that language determines thought: the s p a t i a l metaphors provided by one's langauge determines how one w i l l normally conceptual ize abstract domains. The most thoroughgoing recent treatment of metaphor in everyday language is found in Lakoff and Johnson (1980); they i d e n t i f y the core metaphors that under l i e our t h ink ing about a vast array of domains, and argue that we can understand the domains only by means of these metaphors. The Fundamental Ins igh t that informs a l l t h i s work is t h i s : metaphor is pervasive in everyday discourse and is essen t ia l in our conceptual izat ions of abstract domains. The aim of t h i s paper is to present a computational treatment of metaphor that can accommodate t h i s Fundamental I n s i g h t . 2. Some Frameworks fo r I nves t i ga t i ng Metaphor The e a r l i e s t de ta i led proposal in computational l i n g u i s t i c s f o r handl ing metaphor was that of Russel l (1976). Her proposal concerns abstract uses of verbs of motion and involves l i f t i n g se lec t iona l const ra in ts on the arguments of the verb whi le keeping f i xed the topo log ica l proper t ies of the mot ion, such as source, path and goa l . Thus, to handle " the ship plowed through the sea, " we l i f t the r e s t r i c t i o n on "plow" that the medium be ear th and keep the property tha t the motion is in a subs tan t i a l l y s t r a i g h t l i n e through some medium.