Learn how Workday powers its Optical Character Recognition (OCR) models on the Figure Eight Platform
Workday leverages humans-in-the-loop to train, test, and tune it’s ensemble OCR model to make expenses a breeze
Workday’s stated goal is to put people at the center of enterprise software. They offer a vast array of enterprise cloud products on a single integrated platform, with a specific focus on financial and human resources solutions. They help some of the largest companies in the world run more smoothly by automating and streamlining their payroll, expenses, procurement, and more.
Plenty of us travel for work. And while some people like it a little more than others, the one thing nobody enjoys is filling out their expense report. We’ve all seen the sales team back from a conference, ferreting through a stack of receipts with a look of dread. Companies need their employees to travel and employees need to be reimbursed, but the process of entering every receipt into an archaic expense system is a tedious annoyance no one enjoys.
One of Workday’s cloud products expressly solves for this problem. It’s called Expenses. Essentially, Expenses works a lot like mobile check depositing from a bank does: you snap a picture of a receipt on your phone, the relevant information is captured, and suddenly, your sales team isn’t stuck wasting a few hours filling out an itemized expense report when they should be following up on leads.
Workday handles this with deep learning optical character recognition models, but the challenge here is that receipts are actually a far more OCR problem than checks are. After all, checks are fairly uniform. The signature’s in the bottom right, the amount is written both numerically and in plain English, the account number is down below to the left, and so on. But an airline receipt is much different than a dinner receipt or a hotel receipt. Understanding each means accounting for their differences, recognizing which is which immediately, identifying the relevant fields, and ultimately returning accurate results that result in employees being reimbursed.
All that requires great training data. And for that, Workday turned to Figure Eight.
Workday’s Expenses models has three essential layers. First, there’s a step that identifies fields of interest. Then, there’s a text recognition model that understands what’s in those fields. Lastly, there’s a parser that decides, essentially, what’s important (the phone number of the business is far less vital than the total amount paid, for example).
Figure Eight helps with all three. Our platform enables Workday to annotate the bounding boxes around text for their field of interest model. We power their text recognition algorithms with transcription and output validation. And we help identify which fields are really important by classifying those text strings.
By combining all these Figure Eight jobs with a lot of real-world and synthetic receipt data, Workday has been able to fashion a robust OCR model that understands myriad receipt classes, returning precise results quickly, and making the expense process a lot less painful for their customers.
Creating a model that understands the difference between an airline receipt and a late night run to a fast food place after a long day of selling isn’t as easy at it sounds. It requires a lot of annotated data and an ensemble model that functions well at each stage of the process. The annotations from Figure Eight are a key ingredient in creating a product that makes expense reporting a whole lot easier and a whole lot smarter.