Machine learned document loss recovery

A non-transitory computer readable storage media storing instructions that, when executed by a computer, cause the computer to:

• receive first lossy converted documents, wherein the first lossy converted documents are structured documents in which at least some structure of source documents from which the first lossy converted documents are generated has been lost;
• generate corrected documents for the first lossy converted documents, wherein each of the corrected documents includes edit markers that are based on training input from an operator of a training device and that reflect structure changes in the corrected document relative to a corresponding document of the first lossy converted documents, and wherein the structure changes include removal of at least some of the structure in the corresponding document;
• generate feature vectors for the first lossy converted documents, wherein the feature vectors include structure features of the first lossy converted documents;
• compare first proposed structure edits determined by one or more models based on the feature vectors to structure edits corresponding to the edit markers to determine training data;
• train the one or more models based on the training data;
• subsequent to training the one or more models, apply the one or more models to second lossy converted documents to determine second proposed structure edits, wherein the second lossy converted documents are structured documents in which at least some structure of source documents from which the second lossy converted documents are generated has been lost;
• and
• transform the second lossy converted documents to generate second corrected documents by applying one or more of the second proposed structure edits.

A computer system, comprising:

• memory storing program instructions;
• and
• a processor coupled to the memory and configured to execute the program instructions stored on the memory to cause the processor to:
• receive first lossy converted documents, wherein the first lossy converted documents are structured documents in which at least some structure of source documents from which the first lossy converted documents are generated has been lost;
• generate corrected documents for the first lossy converted documents, wherein each of the corrected documents includes edit markers that are based on training input from an operator of a training device and that reflect structure changes in the corrected document relative to a corresponding document of the first lossy converted documents, and wherein the structure changes include removal of at least some of the structure in the corresponding document;
• generate feature vectors for the first lossy converted documents, wherein the feature vectors include structure features of the first lossy converted documents;
• compare first proposed structure edits determined by one or more models based on the feature vectors to structure edits corresponding to the edit markers to determine training data;
• train one or more models based on the training data;
• subsequent to training the one or more models, apply the one or more models to second lossy converted documents to determine second proposed structure edits, wherein the second lossy converted documents are structured documents in which at least some structure of source documents from which the second lossy converted documents are generated has been lost;
• and
• transform the second lossy converted documents to generate second corrected documents by applying one or more of the second proposed structure edits.