Nunberg (1995) shows that predicate transfer (metonymy) is tied up with all levels of grammar which have an interface to semantics. He explains predicate transfer via a mapping from properties into properties, triggered by a salient transfer function. Various metonymical relations like "owner - things owned" or "producer - product" can be captured in this way. Nunbergian explanations do not, however, easily generalize to metonymies involving representations (maps, statues, models etc.). In these cases, Nunbergs salient transfer function needs to be identified with a representation relation, the properties of which are ill understood. The metonymies discussed in this talk can be taken as pairs, consisting of an NL-expression, and a representation relation defined on things represented and a representing object. The empirical properties of representation relations are discussed with respect to a large corpus of data (videos, transcripts, audiotapes, eye-tracker-studies) based on the construction of a toy-airplane. It is shown that we have different options for building up a theory of metonymy involving representation relations, since we can model the relation involved in different ways. These ways are linked up with (A) the specification of the things represented and (B) the kind of relation used. Concerning (A) there are various choice points for the set of objects presented. Depending upon ones favourite methodology they can be (a) stereotypes, (b) prototypes, (c) Fregean senses, (d) Wittgensteinian states of affairs, (e) situation types or simply (f) objects. Option (f) is more closely investigated. Decisions concerning (B) will, of course, depend upon those taken with respect to (A). We can view representation relations (a) as homomorphic mappings (following Palmer, 1978) or (b) as sequences of morphisms, (c) as suitably defined empirical relations preserving information to some extent, (d) as constraints involving depicted situations and situations depicting or (e) as objects having both, an analogical as well as a Fregean structure. (A) and (B) yield branch points for theories of different explanatory power. Arguments for choices concerning (B) are presented. Stress is laid upon the grain of depictional information and the effects of depicting objects upon objects depicted. It is shown that mappings go in both directions, from objects represented to representing object and vice versa. In the data, metonymies are used for naming parts, planning and testing, fixing of agents' perspectives and agents' coordination. A prototypical case is shown, where a representation relation determines an intrinsic front-back order thereby providing an agent's perspective per default. (Hannes Rieser)