Automatic evaluation of NLG usually involves comparison to human-authored references which are treated as the ground truth. However, these references often fail to adequately capture the range of valid output for an NLG ...