- Andy Schmidt
Move Over OCR – 5 Reasons Why CMR Will Help You Reach IA Aspirations
Optical Character Recognition, OCR, has been around for what feels like forever. It’s become the industry’s most widely accepted tool for data ingestion. Yet, OCR fails spectacularly in meeting the biggest challenge enterprises face – ingestion of all process data, before transportation through RPA. Namely, OCR struggles with accessing unstructured, inferred data and images which comprises 80% of an enterprise’s data.
Previously, these limitations were simply accepted, as there wasn’t anything else available that worked better. Sure, OCR worked if the format was structured and it had a template to follow. OCR is quoted having 80% accuracy, but crucially that accuracy rating only applies to formats and documents it recognises – that means whatever data it doesn’t recognise simply doesn’t get included.
While OCR might work for specific use cases – you might be asking yourself what is OCR capturing? Short answer – not that much. OCR needs a reference to recognise the information presented. Without it, OCR can’t function.
Reason 1: You can’t scale your RPA investment
Most organisations, when starting their RPA journey are advised to start automating the famous “low hanging fruit” processes that include a combination of high potential and low complexity. 'High potential' is driven by volume and estimated time reduction, and 'complexity' by characteristics such as the type of data present. If you have an existing RPA solution, you could be limited, since common OCR-powered RPA solutions primarily handle structured data. Straight-through processing is only achieved with perfectly structured data. That means about 20% of your data is available for digitisation for your automation programme. What happens to the 80% that represents the rest of your data? It’s just not usable for automation. Unless of course, you have available employees to manually key in data for each document. Cognitive Machine Reading (CMR) ingests and digitises 100% of all your data, regardless of format.
Reason 2: You’ll need a template for every single thing.
OCR needs a template to be created each time for ingestion. If there’s no template created, it’s not read. Let’s take a simple use case: invoice processing. Let’s say you have created a basic template for invoices, and it appears your OCR is working fine. But what happens if the invoice format changes? What if the amount or purchase or vendor information is someplace other than where you expect? That will require exception handling later on or manual entry of the invoice automation. Add more suppliers to the equation and these exceptions become standard practice. OK, so it’s time to create more templates. Do you want to spend time creating more templates when the whole point of creating straight-through processing is a touchless process? CMR isn’t template dependant, so variances in zone or format is ingested and digitised, making the data available and usable for business processes.
Reason 3: OCR can’t read semi-structured or unstructured formats.
Semi-structured data, like JSON or XML, OCR can’t read or recognise, let alone digitise. Varied font sizes and special characters present a problem as there is no template to reference. Items like Bill of Lading, purchase orders, insurance policies are a problem for OCR to handle. Plus, OCR errors usually have to be corrected sequentially with the same errors repeatedly edited. Unstructured data formats such as emails, images, handwritten and signatures OCR isn’t equipped to handle. Again, it’s due to the fact there’s no template for it to reference. Non-textual characters, like logos or map symbols, don’t get converted, so they are left invisible until manually examined (assuming you can find the error). CMR ingests all structured, semi-structured and unstructured formats and provides a confidence score of its accuracy. So now more of your data is usable and available.
Reason 4: OCR is lost in translation
Global organisations need to address multi-language data and document formats. If OCR isn’t provided with a template or zone as to what to look for, the information is lost in translation. CMR uses pattern recognition so special characters are digitised. OCR isn’t helpful when it comes figuring out the meaning of data. If a retailer uses OCR to digitise a list and there’s no label associated with it, there’s no context associated with it. So, it either needs to be manually reviewed and the label added, or the data isn’t usable.
Reason 5: OCR’s inability to localise and contextualise
Unless a defined region of interest is created, OCR will attempt to digitise the entire document. But what if you are a reviewing contract, let’s say for LIBOR transition. You know what fields you want to identify, but you just don’t know where within a 100-page contract they might reside. CMR allows you to identify which information is needed within the document. You set the parameters, telling CMR to locate the key information you want and only that. CMR’s Natural Language Modelling (NLM) helps in building models and ontological references (like a heat map) to improve search in the image level which is a whole new way of acquiring data.
Cognitive Machine Reading (CMR) is the only platform that understands, captures and curates every data type. CMR ingests, curates and classifies all data, including unstructured, such as handwritten text, signatures, checkboxes and images and makes it available for downstream policy processing.
This data chasm of non-structured data and unstructured data might not be an issue for your organisation now. But if you want your RPA solution to scale and use all available data, CMR technology may be just the alternative to OCR that you’ve been looking for.