pretrained language models (PLMs) for data-to-text (D2T) generation can use
human-readable data labels such as column headings, keys, or relation names to
generalize to out-of-domain examples. However, the models are well-known in
producing semantically inaccurate outputs if these labe