State-of-the-art machine translation systems often make syntactic errors. Part of the reason is that known models of syntax are either insufficiently expressive or computationally too expensive. To remedy this situation, we introduce Multitext Grammars (MTGs). MTGs generate parallel texts, along with their parallel syntactic structures. We then introduce synchronous parsers. These are efficient algorithms for inferring the kinds of structures generated by MTGs. Finally, we show that syntax-aware translation requires nothing more than a synchronous parser and a trivial post-processor.