I think each key should have a semantic class. For example, "Tel."/" Telephone"/"Tel #", even typo like "Tol #" all belong to a same class. I once interviewed a startup doing this. No expert, but know it is definitely no easy.
the latest version is based on DL. K, V pair labeling is very similar to semantic parsing. If you don't have detailed labels, in my opinion, RL would help in a weakly supervising manner.
【在 w*****r 的大作中提到】 : you sure tesseract's ocr is dl-based?
【在 w*****r 的大作中提到】 : I think each key should have a semantic class. For example, "Tel."/" : Telephone"/"Tel #", even typo like "Tol #" all belong to a same class. : I once interviewed a startup doing this. No expert, but know it is : definitely no easy.