‘We need to establish deep African AI roots...’ – UP’s Prof Vukosi Marivate tells TEDxPretoria how AI can help boost African languages

Artificial intelligence (AI) can be used to connect people and improve our world, but representivity is important in the design of the systems in order to ensure Africa is fairly represented in AI outcomes and processes.

These were some of the thoughts shared by Professor Vukosi Marivate, Associate Professor of Computer Science and ABSA Chair of Data Science at the University of Pretoria (UP), during a recent TEDxPretoria talk. Prof Marivate used three stories to bring to life his journey to promote innovative AI research for African languages, along with some ideas for the audience to ponder.

He recalled how, as a pre-primary learner visiting his school's IBM Reading and Writing Centre, he touched a computer for the first time, and was fascinated by the fact that you could put in some instructions and receive feedback. At that young age, it made him think: What can we actually do with these machines? Later on, in high school, he got to thinking about how you could use data to teach computers to do things; and that idea has shaped his path ever since.

"As a homesick PhD student in the United States, I began spending my evenings researching Tshwane's townships on Wikipedia. Despite the large number of people living in Ga-Rankuwa, for example, where I grew up, there were only a few lines describing the place. This was in contrast to far smaller towns that had more comprehensive entries. So I started adding structure and finding new information about these townships. When I came home and visited a local library, I asked the librarian for any related resources; only for her to come back and hand me a pile of the Wikipedia articles I'd spent the past eight months editing!"

Prof Marivate says this experience taught him a few things: First, it showed that small actions can have a big impact - as this task hadn't really been something he'd planned to do. Second, he learned that Wikipedia acts as ground knowledge for AI tools; and the output doesn't just happen automatically. "If you use the internet and apps, they work on automated decision-making based on what's there, and they've been doing that for decades, so AI is not something new," he said. "The question we need to be asking from this story is: What kind of information is inside these systems? They can be used to connect us and build and improve our worlds, but at the same time, they can be very scary and can be used to drive hate."

This led to his second story, in which he related working on graph mining as a more seasoned PhD student. Using Twitter interactions around the award-winning television series 'Intersexions', on the spread of HIV, it was possible to learn who was influential, and then see how these individuals could help spread the positive message of how to curb the virus. On the flipside was the example of the infamous PR firm Bell Pottinger, which used these same tools in 2017 to run a misinformation and disinformation campaign in South Africa to divide the governing party, and to divide and confuse the citizenry. "We need to understand that if we don't have controls, or don't understand what is really going on, AI just becomes some of us; it can be used for nefarious means, and doesn't represent our ideals, our values, and the things that we want in our societies," Prof Marivate explained. "Representation matters in building these AI systems; if we don't build in what we want, then they don't represent us."

His third story was based on representation of African languages, and he used Setswana as an example, explaining that the current language models would make it appear, at face value, that the people speaking the language don't know anything about astronomy, given the lack of vocabulary available. This is not true at all, he said, and, in addition to knowledge, there are also lots of idioms, stories, and traditions that all have to do with our heavenly bodies.

"In order to correct this, we need individuals studying these things, building these systems, and checking how far they can get in really representing us accurately. We need to make sure that we establish deep African AI roots, so that AI really can be all of us," he concluded.

Watch Prof Marivate's full talk here.