FoodChain ID were going to be using a wide array of AWS services within this partnered project, in order to advance their capabilities and deliver a working end-to-end solution. Services that could very quickly replace old architectures, and turn them into faster, more modernised and well-governed ‘drag and drop’ workflows.

We began with Amazon Textract (an Artificial Intelligence (AI) tool), to automatically extract the text within the PDFs. Textract works cleverly, going beyond simple Optical Character Recognition (OCR), by extracting relationships and defined structures from document types. This helps to reinforce higher accuracy within the labelling results.

Then by setting up Amazon Comprehend, part of the AWS Natural Language Processing (NLP) family of services, we could understand and extract desired insights from text within any document type, perfect for this particular case, where PDF extraction was key.

This is where things get interesting. Rather than simply relying on Comprehend to extract and understand particular values, we used Doccano to drive the retraining data logic, ensuring that there is a corrective learning loop.

Doccano is an open-source annotation tool which provides exceptional text classifications, entity extractions and sequence to sequence translations.

Doccano doubled as both a dataset labeller, and a ‘human’ port of call, for any low confidence scoring from Comprehend. This is useful for the early stages of testing, before deployment, as human operators in both our team and FoodChain ID’s, could easily assess and update the training models – leading to even higher accuracy.