Information Extraction from Receipts with Graph Convolutional Networks

As we have seen before, the Information Extraction step consists mainly of classifying words (tagging), the output can be stored as key-value pairs in a computer-friendly file format (e.g.: JSON). The data extracted can then be efficiently archived, indexed and used for analytics. If we compare OCR to young children training themselves to recognize characters and words, then Information Extraction would be like children learning to make sense of the words. An example of IE would be when you stare at your credit card bill trying to find the amount due and the due date. Suppose you want to build an AI application to do it automatically; OCR could be applied to extract the text from the image, converting pixels into bytes or Unicode characters, and the output would be every single character printed in the bill.