Information Extraction From Document Forms Using Deep Learning

Abstract

Information extraction from printed documents images remains an active re search area. Several methods have been proposed in literature that extract information by utilizing various approaches, e.g., using document geometric or layout information along with various combination of textual attributes. We propose a learning based solution that does not use any layout infor mation and solves this problem using only text blocks contained within the document. We transform the problem into entity relationship mapping and try to find out the probability of a relationship if it is true or not. The method can be used on new documents that are similar in content but can be different in size or layout.