When Africans hear the beat of a talking drum, they comprehend layers of meaning that escape the western ear. Replicating some of the pitches, rhythms and dynamics of human speech, the drum has been used for centuries to relay messages and warnings, convey oral histories, and announce ceremonies and community events.
"In so many communities, the drum is used to share information, or to encode secrets people don't want others outside of their culture to understand," says a computational linguist and artificial intelligence expert Dr. Ife Adebara, a fellow at the Alberta Machine Intelligence Institute and assistant professor at the University of Alberta. "It's a whole communication strategy of its own."
Adebara is one of 25 new faculty members focused on advancing AI research, announced last month at the Upper Bound conference hosted by the Alberta Machine Intelligence Institute (Amii). Thanks to a $30-million investment from the Canadian Institute for Advanced Research through Amii, the goal is to bring together some of the world's best minds in artificial intelligence from a wide range of fields including engineering, environmental science, education and health.
Variations in tone are core features of many of Africa's more than 2,000 languages and dialects, many of which have not yet been committed to writing, she says. "Tone can be the defining factor between two different words with the same spelling; the only difference is tone."
Such aural nuance is the kind of vital data that the large language models of AI have been slow to accommodate, says Adebara, because they are based mostly on text. That means they leave out most of the world's oral languages and dialects.
As linguists remind us, meaning is more than mere words on the page. In all of their rich complexity, languages convey whole systems of knowledge and world views. If you exclude Africa's languages from the AI models, you end up with an incomplete picture of human language in general.
Her research has incorporated more than 500 African languages into the field of natural language processing through data and model development, helping to build what computational linguists call the "corpora" or defining, machine-readable features of language systems. Her broader goal is to make technology accessible to Africans in their indigenous languages by supporting all African languages across both speech and text.
In practical terms, Adebara wants to make AI tools such as search engines, ChatGPT, Google Maps, medical records and technology for agriculture easier for Africans to access and use. Her work also plays a big role in language preservation, since she is also documenting languages for posterity and advocating for policies to keep languages alive.
With extensive training in linguistics and computer science, she also has joint appointments in the U of A's departments of modern languages and cultural studies and media and technology studies, and is an adjunct professor in the Department of Computing Science.
"African languages are spoken by millions of people, but because language policies in most of Africa use foreign languages, there is less data available for indigenous languages," she says.
"The speakers of those languages also feel inferior, because foreign languages have a higher status in society; people would rather learn a foreign language than speak their own. All of this has an impact on how much data is available."
In a previous project funded by the Gates Foundation, Adebara helped collect some 2,500 hours of speech across five African languages to understand the role of tone in computer language models. The researchers found that natural language processing failed to understand tone in isolation, requiring contextual information to give it meaning.
"We know that humans are able to process tone - and attribute meaning - even in the absence of consonants and vowels. These different features are helping us understand not only how human language works, but how human reasoning works."
