Bryant Benson, a leading expert in natural language processing (NLP), enjoys staying on the cutting edge of emerging artificial intelligence (AI) and exploring the newest developments in NLP capabilities.

Can you share the most interesting part of your career journey?

The most interesting part of my career journey is how it started. I came to Deloitte after I finished my PhD in physics, and I had never worked in industry before. I had also never done, or considered doing, any natural language processing. When my first staffing was an NLP project, I threw myself into learning as much as I could in the three weeks I had leading up to joining the project, and as soon as I started learning about NLP I could see that it was going to be an extremely useful skill set to have in this line of work. That was more than four years ago now, and I have been fortunate enough to have stayed staffed on NLP projects ever since. This has allowed me to stay up to date and on the leading edge of those tools.

So when ChatGPT and the other generative text models made such big waves when they came out this past year, I was already using those same technologies and was able to discuss how and why those models work the way they do, as well as help identify potential vulnerabilities and issues for our clients who were considering early adoption of the technology.

What excites you most about working with data and AI?

The opportunity to stay on the leading edge of an emerging technology is really exciting to me. The project I have been staffed on for the last four years has been largely research focused, with our team working to develop tools and bring NLP capabilities into the IRS for broader application. This focus has allowed our team a lot of freedom in exploring the newest developments in NLP capabilities, then using those ideas to create tools and teach the clients how to understand and use them.

In this effort, our team was the first to ever bring transformer-based neural networks (like BERT and GPT) into the IRS, and we were able to start that work only a few months after Google released the model files for their first BERT models.

Describe an interesting project that you have worked on.

One of the tasks we worked on at the IRS was to extract and validate information on a form that was mostly paper filed. This information was needed because a law had changed, and the IRS was at risk of having to repay more than $2 billion in collected interest if they could not confirm the information on this form for relevant years. Our team developed a hybrid solution that used text classification to identify the necessary information from the forms and optical character recognition (OCR) was used to extract the required information. We used text cleaning to correct minor OCR mistakes in extracted dollar values.

The clients originally estimated that it would take 20 full-time employees for a whole year to read through the document packets, find the forms, then transcribe and validate the information. Our team of five built out this solution in six weeks.