Systems, methods, and programs embodied in a computer readable medium are provided for index extraction. Stored in a database are ground truth documents that are organized according to a plurality of classifications, each classification having a group of predefined indices. A document to be indexed is classified by drawing an association between the document and one of the classifications. An attempt is made to extract from the document at least a subset of the group of predefined indices associated with the one of the classifications. Upon a failure to extract the subset of the group of predefined indices, attempts are made to find and correct at least one text recognition error in the document based upon a salient dictionary associated with the one of the classifications.