0
0
Machine learning has significantly increased document processing efficiency in the mortgage industry. However, the effectiveness of any technology depends on document quality and a vendor’s ability to train machine learning models on a broad set of document types. As such, it’s wise to be skeptical of broad accuracy claims being made across the industry from start-ups and seasoned vendors alike.
To help remove the mystery and confusion surrounding accuracy I am sharing five important questions you should ask of the vendors you’re evaluating for mortgage document processing.
First, to level set. What is accuracy?
Accuracy is how often an automated process, such as automated document recognition or automated data extraction, is correct in its resulting outcomes. For example, if automated document recognition is presented with 100 documents, and it is correct in its resulting outcomes on 89 of those, it is then 89% accurate.
Now, onto our five questions.
- Are accuracy numbers based on narrow set of documents?
The doc type repository that LoanLogics recognizes through its automated document recognition (ADR) technology has climbed towards nearly 500 unique document types! Of those documents, the industry’s most critical doc types, see about a 95% accuracy rate using automation alone!
2. How long does it take to “train” a new document type?
After processing 987 million unique documents and 3.02 billon pages over the last 10+ years, suffice it to say our technology has a lot of training samples. This provides immediate value to LoanLogics clients on day one. And the good news is, it is only getting better with the continuous training our machine-learning models receive that helps to improve the speed and accuracy of document recognition.
Other vendors with less experience, may be using client documents as the initial training set for their solutions. The system may be able to be easily ‘trained’ by ingesting the document but how accurate will it be having only processed fewer versions of it.
This infographic showcases the full robustness of LoanLogics doc processing experience and automation technology.
3. How is accuracy impacted by poor document quality? Were the documents first generation or copies?
If we had perfect pristine documents, the mortgage industry would be a lot simpler. But the fact of the matter is, as lenders, servicers, insurers, aggregators and vendors we all continually receive coffee-stained, poorly scanned and undiscernible documents. If a doc processing vendor’s technology fails to deliver confidence in its automated results, it is important that they have a way to overcome this. LoanLogics offers a well-integrated exception management human workflow, in both its service and on demand technology, that can intercede and provide the clarity that the automation was unable to provide
4. Can accuracy be achieved across both unstructured or structured document types?
For LoanLogics, yes. Our doc processing technology automation handles both structured and unstructured doc types, their vast variation within a document type, all with great accuracy. By going beyond template-based approaches, machine learning uses algorithms that can recognize patterns to cluster & classify. As already stated, the more samples used to train machine learning, the better it gets.
5. Is the technology combined with human assisted workflow to achieve accuracy levels?
From a LoanLogics perspective, IDEA Services adds that human assisted workflow to achieve our 99% data accuracy benchmark. There are clients of ours and across the industry who require the utmost data accuracy for their specific application use case. By combining human operators with our “Automate First’ strategy we are able to deliver optimal results with a more strategic use of humans in our document processing workflow when required.
I hope you found these thought provoking questions of interest and invite you to learn more about our capabilities by either visiting our website or reaching out at loanlogicsinfo@loanlogics.com. This recent LoanLogics article “Removing Costs With Confidence,” featured in DS News earlier this year takes the conversation one step further discussing the importance of a technology’s ability to use confidence scoring to further refine the output of a mortgage organization’s doc processing to their individual use cases and needs.
