Enhancing the quality and accuracy of OCR results

The Need

Text mining tools powered by natural language processing (NLP) have become indispensable for modern businesses. The quality of NLP solutions heavily depends on the accuracy of the optical character recognition (OCR) engine used. OCR is a technology based on sophisticated computer vision algorithms for recognizing characters from printed books, handwritten papers or images in all possible fonts, sizes, and orientations. With this technology, companies can quickly transform document images into electronic, machine-readable text - the input for many downstream processes, among which text mining solutions.

OCR accuracy has improved over the years and is considered by many as a solved problem. Nevertheless, errors still occur in practical application, which, if left undetected, can skew results – especially where machine learning NLP algorithms are involved. Sources with tiny font sizes, blurred copies or colored paper can trip up OCR algorithms. The resulting electronic text errors are easy to detect for human readers, who are generally able to infer something is wrong. NLP algorithms, however, could completely misinterpret the text, leading to failure of the text mining tool, which requires manual work finding and correcting the failed conversion

Our solution: DocuMend

Deloitte DocuMend identifies the issues by assessing the accuracy of OCR layers both at a document and individual word level. Users may choose among several OCR engines and set the confidence threshold for OCR quality. The accuracy assessment is overlayed onto the original text in the PDF as a sort of textual heat map – ranking the OCR-processed words from lowest to highest confidence in OCR quality.

Up until now, quality control of OCR required humans to proof-read, finding errors either through context or intuition. DocuMend takes advantage of multiple OCR engines to cross-validate, highlighting discrepancies between the engines as likely sources of error. To correct the identified errors, users navigate through the textual heat-map document or work directly from a list where identified errors are ranked from most to least certain, down to the user-specified threshold.

Advantages/Benefits

• Automated assessment of OCR quality across multi-page documents
• Visual feedback via word-level confidence heat-maps
• Intuitive, no-code graphical interface allows business users to quickly navigate documents and gain confidence in the quality of their OCR layers
• Direct user-interaction to correct errors on the spot

Example Use Cases

• Improve quality of CV screeners for AI-supported HR recruiting processes
•  Improve reliability of fit-and-proper screening for banks with regulators
• Increase confidence in digital document processing adoption within the business
• Lay a strong foundation for countless AI applications making use of text-mining / NLP

Back to main page