Natural language generation (NLG) systems rely on corpora for both hand-crafted approaches in a traditional NLG architecture and for statistical end-to-end (learned) generation systems. Limitations in existing resources, however, make it difficult to …