This paper describes an attempt to combine the advantages of both example-based translation and stochastic translation methods in an attempt to develop a method for inferring symbolic transfer functions from a bilingual corpus. By formalizing the translation process and applying standard optimization techniques, a system can be developed that will identify grammatical categories and produce coherent transfer functions between languages. The validity of this approach is demonstrated in a prototype system that can learn transfer functions between English, French, and Urdu. Contains 13 references. (Author)