Property mapping is a fundamental component of ontology matching, and yet there is little support that goes beyond the identification of single property matches. Real data often requires some degree of composition, trivially exemplified by the mapping of "first name" and "last name" to "full name" on one end, to complex matchings, such as parsing and pairing symbol/digit strings to SSN numbers, at the other end of the spectrum. In this paper, we propose a two-phase instance-based technique for complex datatype property matching. Phase 1 computes the Estimate Mutual Information matrix of the property values to (1) find simple, 1:1 matches, and (2) compute a list of possible complex matches. Phase 2 applies Genetic Programming to the much reduced search space of candidate matches to find complex matches. We conclude with experimental results that illustrate how the technique works. Furthermore, we show that the proposed technique greatly improves results over those obtained if the Estimate Mutual Information matrix or the Genetic Programming techniques were to be used independently. The final publication is available at Springer via http://dx.doi.org/10.1007/978-3-642-40285-2_18.