Using a Feature-Based Lexicalised Tree Adjoining Grammar (FB-LTAG), we present an approach for generating pairs of sentences that are related by a syntactic transformation and we apply this approach to create language learning exercises. We argue that the derivation trees of an FB-LTAG provide a good level of representation for capturing syntactic transformations. We relate our approach to previous work on sentence reformulation, question generation and grammar exercise generation. We evaluate precision and linguistic coverage. And we demonstrate the genericity of the proposal by applying it to a range of transformations including the Passive/Active transformation, the pronominalisation of an NP, the assertion / yes-no question relation and the assertion / wh-question transformation.