Natural language generators (NLGs) for task-oriented dialogue typically take\na meaning representation (MR) as input. They are trained end-to-end with a\ncorpus of MR/utterance pairs, where the MRs cover a specific set of dialogue\nacts and domain attributes. Creation of such datasets is labor-intensive and\ntime-consuming. Therefore, dialogue systems for new domain ontologies would\nbenefit from using data for pre-existing ontologies. Here we explore, for the\nfirst time, whether it is possible to train an NLG for a new larger ontology\nusing existing training sets for the restaurant domain, where each set is based\non a different ontology. We create a new, larger combined ontology, and then\ntrain an NLG to produce utterances covering it. For example, if one dataset has\nattributes for family-friendly and rating information, and the other has\nattributes for decor and service, our aim is an NLG for the combined ontology\nthat can produce utterances that realize values for family-friendly, rating,\ndecor and service. Initial experiments with a baseline neural\nsequence-to-sequence model show that this task is surprisingly challenging. We\nthen develop a novel self-training method that identifies (errorful) model\noutputs, automatically constructs a corrected MR input to form a new (MR,\nutterance) training pair, and then repeatedly adds these new instances back\ninto the training data. We then test the resulting model on a new test set. The\nresult is a self-trained model whose performance is an absolute 75.4%\nimprovement over the baseline model. We also report a human qualitative\nevaluation of the final model showing that it achieves high naturalness,\nsemantic coherence and grammaticality\n