A Statistical Approach to the Processing of Metonymy

This paper describes a statistical approach to tile interpretat ion of metonymy. A metonymy is received as an input, then its possible interp retations are ranked by al)t)lying ~ statistical measure. The method has been tested experimentally. It; correctly interpreted 53 out of 75 metonymies in Jat)anese. 1 I n t r o d u c t i o n Metonymy is a figure of st)eech in which tile name of one thing is substi tuted for that of something to which it is related. The czplicit tc.~m is ' the name of one thing' and the implicit t;c~"m is ' the name of something to which it; is related'. A typical examt)le of m(;tonymy is He read Shal(esl)eare. (1) 'Slmkesl)(~are' is substitut(~d for ' the works of Shakespeare'. 'Shakest)eare' is the explicit term and 'works' is the implicit term. Metonymy is pervasive in natural language. The correc~ t rea tment of lnetonylny is vital tbr natural language l)rocessing api)lications , es1)ecially for machine translation (Kamei and Wakao, 19!)2; Fass, 1997). A metonymy may be aecel)table in a source language but unaccet)table in a target language. For example, a direct translation of 'he read Mao', which is acceptable in English an(1 Japanese, is comt)letely unacceptal)le in Chinese (Kamei and Wakao, 1992). In such cases, the machine trmlslation system has to interl)ret metonynfies to generate acceptable translations. Previous approaches to processing lnetonymy have used hand-constructed ontologies or semant ic networks (.]?ass, 1988; Iverson and Hehnreich, 1992; B(maud et al., 1996; Fass, 1997). 1 1As for metal)her l)rocessing, I,'errari (1996) used t;exSuch al)t)roaches are restricted by the knowledge bases they use, and may only be applicable to domain-specific tasks because the construction of large knowledge bases could be very d i f ficult. The method outlined in this I)apcr, on the other hand, uses cortms statistics to interpret metonymy, so that ~ variety of metonynfies can be handled without using hand-constructed knowledge bases. The method is quite t)romising as shown by the exl)erimental results given in section 5. 2 R e c o g n i t i o n a n d I n t e r p r e t a t i o n Two main steps, recognition and i'ntc.'q~vcration, are involved in the processing of metonyn~y (Fass, 1.!)97). in tile recognition st;el), metonylnic exl)ressions are labeled. 1111 the intel'l)r(:tation st;el) , the meanings of those ext)ressions me int, eri)reted. Sentence (1), for examl)le, is first recognized as a me tonymy an(t ~Shakespeare' is identified as the explicit term. 't'he interpretation 'works' is selected as an implicit term and 'Shakespeare' is replaced 1)y ' the works of Shakespeare'. A conq)rehensive survey by Fass (]997) shows that the most COllllllOll metho(1 of recognizing metonymies is by selection-restriction violations. Whether or not statistical approaches can recognize metonymy as well as the selectionrestriction violation method is an interesting question. Our concern here, however, is the interpretation of metonymy, so we leave that question for a future work. In interpretation, an implicit term (or terms) that is (are) related to the explicit term is (are) selected. The method described in this paper uses corpus st~tistics for interpretation. tual clues obtained through corl)us mmlysis tor detecting metal)lmrs.