The Department of Health and Human Services (HHS) has piloted the use of NLP to process public comments on new regulations, which can require over 1,000 hours just to categorize for a single proposed rule. The tool was able to meet quality requirements and improve staff satisfaction, allowing one agency to demonstrate millions in cost savings.
The low-hanging fruit here may well lie with the agency help desk, where AI can be trained on the FAQs. NLP could route calls effectively, easing the burden on help desk staff, and could even help to resolve queries that are purely informational. In mature contact centers, Accenture has found that costs can be reduced by 30 percent with higher customer satisfaction through expanded use of more intelligent virtual assistants.
Some agencies already are moving in this direction. U.S. Citizenship and Immigration Services (USCIS) for instance has introduced Emma, a voice-powered personal assistant that can understand and speak both Spanish and English. Other agencies are looking to Emma as a model for what may be possible on the citizen-service side, with natural language enabling organic conversations and helping to fulfill routine requests with little to no human intervention.
As our population continues to age, finding new ways to enable the elderly to lead productive, independent lives will grow in importance. An Accenture pilot in the United Kingdom used Amazon Echo devices to empower caregivers to provide more virtual care and support. And working with the UK’s National Theatre, we developed a device using NLP for real-time audio captioning for those with loss of hearing.
While these use cases focus on the spoken word, government also may have much to gain from the ability to both analyze and generate text.
For example, Accenture is working with government agencies responsible for processing benefits claims for citizens. A key challenge is the inordinate volume of documentation requiring manual review to ensure information is correct and consistent. NLP can do that, “reading” documents at computer speed to ensure citizens are eligible for services or benefits. The machines won’t make the final call—no one likes to think of AI rejecting a health care claim!—but they can flag suspicious entries, inconsistencies or apparent discrepancies. Rather than wade through 100 pages of detail, a human reviewer can skip to the two pages where things look sticky, or the computer can automatically confirm that claims are supported by the text.
NLP is also a core technology for link analysis, which allows analysts and researchers to make network correlations across data, whether it be medical research, search engine optimization or criminal investigations. In this context, NLP technologies like latent semantic indexing can play an important role in concept matching, eDiscovery and disambiguation, allowing conceptual relationships to be identified even when not readily apparent.
In the near future, some expect NLP will become adept not just at reading government documents, but at writing them, too. Many agencies are tasked with generating reports, typically based on tables of data. What if the machines could read the data and formulate human-sounding sentences and paragraphs?
Another scenario that’s easy to envision centers upon document retention. Various agencies are charged with keeping documents for certain lengths of time, in response to Freedom of Information Act requirements and other regulations. Typically, documents are marked for retention by a manual process: That’s time consuming, tedious and error prone. A better idea would be to have the machine read and classify documents, automatically filing them according to authorized release dates.
In each case, the sweet spot for NLP lies somewhere at the intersection of simplicity and repeatability. AI shines brightest when it helps to automate routine tasks, where the level of complexity is fairly low and where the high frequency gives the AI the opportunity to learn from many repeated instances.
How exactly could government put NLP to work, and who would be in charge of that? It’s worth a deeper dive.