Services Australia is trialling an optical character recognition (OCR) solution to greatly reduce the amount of time it takes Centrelink paperwork to reach caseworkers for assessment.
First introduced at the height of the coronavirus pandemic, the Capgemini-built solution has been pivotal to accelerate claims processing during a period of increased demand for welfare support.
It uses artificial intelligence and machine learning to automatically check whether information contained in forms lodged by welfare applicants and recipients is accurate and complete.
Some 25,000 documents currently pass through the agency’s document lodgement service each day, meaning it can be days - if not weeks - before caseworkers can assess claims.
Capgemini CEO and AN/Z managing director Olaf Pietschner told iTnews that the solution was introduced just in time for the rapid increase in claims that resulted from Covid-19.
“We started working on this in August 2019, so before Covid, and we built a trial - that’s still undergoing testing - and it’s contained to a particular area of [Centrelink] documents,” he said.
“We had this set up, deployed and tested when the amount of digital transactions that citizens were looking to interact with Services Australia [increased].
“And its then been a solution that helped deal with increased workloads and increased digital interactions with citizens.”
Capgemini public services vice president Lysandra Schmutter said that by automating document checking, the solution has been able to “rapidly improve” processing times.
“We are able to rapidly improve the time it takes from when a citizen lodges a document to when it's placed in the hands of a caseworker,” she told iTnews.
Asked to quantify the improvement, Schmutter said “we’re working with Services Australia through what that really means in terms of quantifying those outcomes”.
The project is separate to work Capgemini has performed as part of the third tranche of Services Australia’s seven-year, billion-dollar welfare payments infrastructure transformation program.
Pietschner said the first phase of the project involved delivering the “containerised machine learning model and application to identify a document against predefined categories”.
He said the model had an accuracy rating of “more than 95 percent” for the documents that were processed, giving the agency the confidence the model would meet its needs and ease workload.
The API-based solution itself is coded in Python and utilises “cutting edge approaches … for OCR” that were developed by Capgemini during the engagement with Services Australia.
As part of this, the team “developed a progressive scale expansion neural network (PSENet) working in tandem with regression-based neural network (RotNet) and Tesseract OCR engines to produce accurate and usable OCR”.
“It’s really using an advanced neural language processing model and understanding the importance and the differences of words between documents to define the different types,” Pietschner said.
“And then that information is used to do this auto classification and validate that the citizens have uploaded the correct set of documents.”
Capgemini plans to continue testing and proving the accuracy of the solution until the completion of the trial, at which point it hopes to roll out the solution more extensively across Services Australia.
“Certainly the solution we developed for this lends itself to a much broader application, really leveraging machine learning, natural language processing and neural networks across a much broader range of applications,” Pietschner said.
Other areas within Services Australia that could benefit from the technology include Child Support and Medicare, though the solution could also be extended to agencies across the government more broadly.
“We’re really excited for what this solution can actually prove to government and assist with government, and we think that there’s multiple uses for this type of solution,” Schmutter added
“If you think quite broadly across all government agencies, where documents of scale are lodged for any claims or any type of request, we feel that this absolutely something that could be used and harnessed.”