In order to improve the usability, flexibility and agility of model-driven engineering (MDE), various approaches have been adopted to facilitate the creation and adaptation of MDE tools by end-user software practitioners, and to reduce the level of MDE skills necessary for such tasks. Two important techniques are the use of examples to specify MDE tools such as model transformations, and the use of specifications based on the concrete syntax of the software languages being processed by transformations, instead of the language metamodels. In this paper we combine these two techniques with symbolic machine learning, in order to derive language-to-language transformations from sets of examples expressed using concrete syntax. The approach is demonstrated on a program abstraction task (i.e., reverse engineering), a DSL-to-code code generation task, and a refactoring task.