Document Understanding using LSTM (Long Short Term Memory) Neural Networks

Abstract

To endure through the throat cutting contention, business industry and researchers are compelled to take intelligent and proficient decisions. This gives rise to the requirement of an automated solution powered by cutting edge technology that can extract the information in the blink of an eye and concoct unprocessed scanned image data, emanating in resourceful and perspicacious information at one’s disposal. The competency can be consummated through evaluating the scanned document in a certain manner. Nonetheless, to extract bibliographic data from the scanned document image, there is a necessity of understanding of complex semantics of structured textual and visual data, which is precisely what majority of the existing automated solutions curtail. Therefore, our ambition was to develop the Document Understanding system that will be able to analyze the document scanned image data considering the current business and research requirements of an organization. Powered by the highly acclaimed Long Short Term Memory (LSTM) neural networks and complemented by its rich and user friendly interface on top, our system is a dynamic bibliographic extracting solution. Our solution will save your precious time. This thesis presents a document understanding system for scanned images of medical journals articles documents. Using Long Short Term Memory(LSTM) neural network, the bibliographic data that includes titles, authors, affiliation and abstract is extracted conveniently. Results show phenomenal agreement with theoretical predictions and significant improvements over previous efforts in this domain. The work presented here has abstruse and sagacious implications for future studies of logical analysis of layouts of documents and may one day help solve many such existing problems.