Solve the African language problem for inclusive AI development

ITU News

Artificial intelligence (AI) technology can help leapfrog major global development challenges – but not until the African language problem has been solved.

Opening opportunities to participate in the digital economy depends on access to technology, and if you cannot understand English or one of the handful of other languages that dominate tech, you miss out.

This was the message at the heart of an AI for Good Global Summit keynote by Pelonomi Moila, CEO of Lelapa AI, a socially grounded research and product lab developing language technology for African nations and communities.

The tremendous potential of AI to drive development must be weighed against the risk of deepening existing inequalities. This is particularly clear in the context of Africa, with large numbers of the population living at great distances from essential services and products such as medicine, education, finance, and government access the Internet by mobile phone.

Development challenges can be addressed at scale with lower costs and greater reach by providing products, tools, and services – from agriculture to finance or teaching – via mobile.

But the opportunity is limited, because the lingua franca of tech and telecom services is English – which much of Africa does not speak.

‘’If you don’t speak one of the major European languages, you cannot access digital products and services which are not multilingual – from chatbots to voice instructions or government services,” said Moila. Consequently, when your access to education is mainly online, “You have to learn English before you can learn maths or science.”

Watch the keynote speech at the AI for Good Global Summit.

Digital exclusion of the Global South

Big tech’s Large Language Models (LLMs) may have solved the issue in the northern hemisphere. But big tech won’t solve the issue in Africa, where there is not enough commercially available computing power, little funding for local organizations focused on AI African content, and very limited data sets.

Of the top 34 languages used on the Internet globally, not one single one is African. This partly reflects historic prevalence of the continent’s oral cultures, meaning there are no huge bodies of text to be scraped for information.

Open source LLM development is not the answer either, at least until there is sufficient access to computing for local language communities to generate and licence their own data, along with sufficient content using and creating information in local languages.

The solution may lie with the approach adopted by the African AI Network, a collection of tech communities, labs and grassroots movements engaged in collaborative research and development of African language solutions. Spread across eight countries, the group is initially working with eight languages – with the potential to connect 520 million speakers across the continent to the global digital society.

The aim, Moila said, is two-fold:

• Firstly, “Building technology for Africans to communicate with technology in languages that allow access to products, services, and communication.”
• And secondly, “Building technology to bring people together, not technology to divide the haves and have-nots.”

Less is more

Instead of data-hungry LLMs, which are expensive to develop, run and train, the initiative favours low-data, low-compute options optimized for the constraints on the ground.

Smaller language models working in combination can cost less and perform better, producing fewer hallucinations (non-sensical or inaccurate LLM output) as they are focused on smaller domains or specific problems, she added.

Transcription, conversational and text analysis tools function well using much smaller volumes of data – and those developed by the African AI Network are often the only models available for the languages in question.

Moila called for international support in building more of those essential smaller models for content creation, translation, and digital or AI applications across Africa. Compared to big LLMs for major world languages, the models would use far less resources and contribute significantly less to the global climate crisis, she asserted.

They would be trained on cues from African social systems, she added.

Support needs to be directed to local communities on the ground, bringing to the table the people who understand what is needed, what will work and how to measure success – “building with us, not for us,” she added.

“Solve the language problem with less, not more. Join us in the quest for equity around language solutions.”

Solving the “African language problem,” she concluded, will be key to achieving the UN Sustainable Development Goals.

Watch the interview here.

Header image credit: ITU