One of the main obstacles to the development of any text generation system is the difficulty of creating the knowledge base from which texts are generated. In particular, a system which performs sentence generation "from scratch" requires a great deal of information about how concepts in its knowledge base are expressed syntactically. Unsurprisingly, it has proven very difficult to build tools which acquire such information from people with no specialist expertise in linguistics and semantics. However, there is another method for acquiring such information which has been largely overlooked: if we want to build a knowledge base that allows full sentence generation, all we need is a bidirectional sentence processing system in which a single declarative grammar supports both sentence generation and sentence interpretation. With such a system, authoring new facts in the knowlege base can be achieved very naturally in the context of a natural language dialogue.