Organizing African talent to move humanity forward: Language technology for Africa

Pelonomi Moiloa Advocates for AI Inclusivity in Africa, Addressing the Language Barrier Challenge

In a compelling and thought-provoking speech, Pelonomi Moiloa, co-founder of Leelapa AI, highlighted the transformative potential of artificial intelligence (AI) in Africa, while also drawing attention to the significant language barriers that impede progress. Moiloa engaged her audience with an interactive demonstration to illustrate the frequent misunderstandings between AI and humans, particularly when it comes to accent and language recognition, showcasing the inconvenience and exclusion that can result from such technological shortcomings.

Central to Moiloa’s address was the “African language problem,” which she described as the failure of AI to support the vast array of African languages. This deficiency not only hinders non-English speakers from engaging with the digital economy but also limits their access to crucial services, thereby perpetuating inequality. Moiloa argued that while AI has the capacity to address developmental challenges in Africa, this potential is stifled by the existing language barrier.

Moiloa underscored the critical role of mobile technology in Africa, where mobile phones are the predominant means of accessing the internet. She pointed out the opportunity to expand the delivery of essential services such as healthcare, education, and financial services via mobile platforms. She also noted the advanced state of mobile money systems in Africa, which presents a unique opportunity to facilitate transactions and service access.

Introducing the African AI NLP network, Moiloa showcased a collaboration of leading labs, communities, and organisations, including Leelapa AI, which are dedicated to developing AI technologies tailored to African languages. These initiatives aim to create AI models that enable Africans to communicate digitally in their native tongues, thereby narrowing the digital divide.

Despite the promise of these initiatives, Moiloa detailed the significant challenges encountered in developing AI for African languages. These include a lack of data, as African languages are often poorly represented on the internet; limited computing resources, with only one commercially available data centre with GPUs on the continent; and a dearth of funding for AI startups in Africa, with a mere fraction receiving venture capital support.

Moiloa’s call to action was a plea for collaboration and support from the global community. She emphasised the importance of inclusive development, advocating for the global community to build with Africans, ensuring they are part of the conversation. She also highlighted the limitations of relying on open-source models without also providing compute resources and supporting local data generation and ownership. Furthermore, Moiloa advocated for the development of smaller, more resource-efficient AI models that are better suited to the constraints of the African context.

In her conclusion, Moiloa reiterated that addressing the African language problem is crucial for achieving the Sustainable Development Goals (SDGs) and that effective solutions must include African voices and expertise. She invited the audience to join the mission to ensure that AI technology meets the diverse linguistic needs of Africans, fostering greater participation in the digital economy and contributing to sustainable development.

Throughout her speech, Moiloa conveyed a message of urgency blended with optimism, positioning Africa at the forefront of a technological revolution that could be inclusive and empowering if the right mindset and resources are applied. She celebrated the creativity and resilience of African innovators and emphasised the value of harnessing local knowledge and talent to forge solutions that are both impactful and sustainable.

Pelonomi Moiloa:
Oh, OK. So to start, just to set the context a little bit better, I’m going to ask you to bear with me. And please humor me and lift your arm in the air. I’m not going to make you hold it there for too long, but I would appreciate if you would take part in this exercise. Great. So we maybe have about 40% of the room with arms in the air. Please put your hand down if you’ve ever had to change your voice, either the tone of your voice or your accent, in order for AI technology to understand what you’re saying. OK. We’re down to about maybe 10% left. We have a couple that we can still work with. Please put your hand down if you’ve ever had AI say your name in a way that completely butchers it beyond recognition, your name or the name of a friend. A few more hands have gone down. And please put your hand down if you’ve ever texted somebody in your family or in your life, and auto-correct or something similar has made you send a really embarrassing message. I think all the hands are now finally down. Cool. Great. So for many of us, engaging with technology that doesn’t understand and that we can’t communicate with is a minor convenience. But for many more people all around the world, and in the majority world, technology not understanding you can be the difference between having access to opportunities or missing out on opportunities to participate in the digital economy. And what this means is that we have a very severe problem in terms of equity and equality access to technology. But at the same time, I’m a desperate optimist, it’s also a really incredible development opportunity to leapfrog developmental challenges. So on the African continent, we have large populations of people that are situated quite far from all kinds of essential products and services. This can include medical care, education, financial institutions, et cetera. And people can take up to days in order to travel to be able to access these essential services and products. And on the African continent, internet is mostly accessed via mobile phone. It’s not always a smartphone like you might be taking pictures of the screen with now. It is often a feature phone. But it does mean that the mobile phone and mobile penetration on the African continent has given us an incredible opportunity to increase the access to products and services via mobile. And this means we’re able to scale up access to these products and services, but at a lower cost. And this means for things like the SDGs that we have opportunities to approach developmental challenges with lower budgets, but with greater reach. So in the area of zero hunger, we’re able to give access to farmers to information that could help them increase their crop yields or navigate through risks and challenges posed by weather events or extreme other indicators. There’s an opportunity for the commercial sphere in terms of being able to use mobile money, which, by the way, on the African continent, our mobile money is far advanced compared to the rest of the world. But it gives us an opportunity to use that mobile money to be able to pay for essential products and services like never before. In terms of the medical sphere, we could potentially use this mobile access in order to give people an opportunity to make bookings at medical clinics that are nearby, ensuring that people are there and available to tend to the sick and the unhealthy. We’re able to ensure that there is a medical worker available at that time to tend to that person to manage loads of medical influx. And at the same time, we’re also able to use this as an opportunity for local organizations, local economic ecosystems, and startups to build applications that are able to extend the reach of those products and services into the greater world and the digital economy. We’re also able to potentially use this technology in order to progress the level of education on the African continent. We’re able to give them applications and tools to be able to consume the information on the internet so that they don’t have to learn English before they’re able to learn math and science. But as you may have noticed, this opportunity is severely limited by the African language problem. And at this point, you might say, well, what is the African language problem? And the African language problem is briefly framed by the fact that the lingua franca of technology is English. This means that if you’re living on the African continent, you don’t speak English, or you don’t speak Portuguese, or French, or some other language that technology knows well. And the person who’s developed the application through which you are trying to access digital products and services has not put in a concerted effort in order to make that service multilingual due to resource constraints or budgetary concerns. It means you have no access. And this means that every chatbot or voice instructions or calls that are monitored for quality assurance purposes, being able to access government surveys, et cetera, all through your mobile phone, over the internet, you do not have access to if you do not speak the language or technology. And so AI technology can really help us sleep for our developmental challenges, but it’s just not possible until we solve this language problem. And this is the part where you say to me, but isn’t language problem solved already by big tech? I mean, there’s been such an uptick and excitement around the hype and opportunity of large language models since February of last year. I thought this problem was solved. Well, it’s not solved. There is a very long tale of the majority of the world’s languages that are not covered by this advanced technology, and we’re being left behind. And it’s also not being solved by big tech because they’re not the right people to do it, but we are. And who are we? We are the African AI network, more specifically the African Language Solution. The African AI NLP network consists of top labs on the continent, thriving communities, such as the Deep Learning Endeavor, and the Masakaneti grassroots research movement, which isn’t only just pioneering the way in terms of NLP, low resource research, but also in terms of collaborative research and how to do that, but also with successful organizations. Leelapa AI is one such organization, and the organization of which I am a co-founder. And we build the technology for Africans to communicate digitally in our languages. So that technology can hear us when we speak to it. We don’t have to change our accent or the language that we speak in order to be understood. It means that technology can speak to us without us having to learn a new language in order to communicate, and so that we can have conversations with technology in ways that can help us gain access to products and services. And Leelapa AI, in its essence, is really an initiative to try and change the way that we build technology, such that we’re able to bring people together rather than divide the haves from the have-nots. We’re in this incredible situation where we have this open suite of African problems, this top African AI talent network that we’ve helped to build, and some seasoned African AI founders. And we really have these two initiatives that we’re driving quite hard around becoming a leading research lab, not just for the African continent, but also teaching the world how to develop AI technology in the language space with more resource efficiency, but also to commercialize research and development on the ground. Robots moving around and doing really cool things is quite exciting, and it is new. But what problem is it solving? And that is the question we ask ourselves every day. What problem are we solving? We’ve been around for around 15 months, and in that time, we’ve managed to hold on to 20 permanent staff within the African context, preventing them from contributing to the brain drain issue of Africans going into other regions for other opportunities and spending their time and effort in those places. We are spread across eight African countries and collectively have contributed to over 20 publications. We have built 19 models, and I’ll speak a little bit more about that a little bit later. And of those models, we have created three products, one of which is our Vula Vula API. It enables five features, so five mechanisms of human communication. That is speech, hearing, basically conversation, being able to analyze and understand text. And we are focusing on just eight languages for now. But those eight languages help to enable us in enabling 520 million people across the African continent and giving them the option to participate in the digital economy through language. And we are optimized for these two key things, resource-constrained environment, which is the opposite of how technology is currently being developed in the world. We deal with low data and low compute. And we also focus on the applied research domain. So this is the area where we constantly ask ourselves, technology is great, technology is cool, but what problem is it solving? And so I’ll use the case of a call center agent in order to illustrate this, because often we just get the question, why can’t you use an LLM? So let’s just say, for example purposes, that we could use an LLM because LLMs worked on our languages. Typically speaking, you’re looking at an estimated $100 million to be able to train such a machine. You’re looking at needing 45 terabytes of data somewhere around there. And every time you serve the model or respond to a request, you are needing to spend around $0.01 in order to do so. And then you get your response. In our case, the L in LLM stands for little rather than large. And instead, we take the approach of ensembling smaller, more powerful models together in order to achieve the same result. So we can stick together an intent model, an NER model, a search model, map it to a response, and we’re able to significantly reduce the amount of money used to train that model. We need far less data, and it means that we can serve it at a much cheaper price. And so with smaller models, you really are able to have an advantage in terms of being able to do what you need to do. You’re able to get better performance without the threat of hallucination because you’re focused in on a particular domain and a particular problem that you’re trying to solve. You require less resources. This is both from a data perspective and a computer perspective. And you make these models more accessible because they’re cheaper to train and they’re cheaper to use. And so just an example of how this reflects on three of the features that we’ve come to build. Transcription is an enabling conversion of audio to text. Here, instead of the typical 100s of thousands of hours to train a model, we’ve used 150 hours. The model that we produce is around five times smaller than the typical commercially available model for this, which is Whisper. And we’re able to decrease the word error rate by 14% average across the languages that we’re building for. In the conversational domain, we’re not at the generative stage yet. But in terms of enabling conversational chat, we’re using few-shot type learning. So this means just a few examples of data in order to be able to tend to the solution of a problem. This model is around 120 times smaller than a generative LLM. And if we’re comparing to what exists on the market for our languages currently, those models don’t exist. So we don’t have a performance measure for that yet and how much better it’s doing than the state of the art. And the same can be said for the analysis domain. So this is the opportunity of gaining insight into the text that you create through things like sentiment analysis and such. Here, we use a model called an XLMR. It is pre-trained, so there is a bunch of data that has gone into that. But on top of it, we’ve only used 9,000 tokens to help refine and fine-tune it. This model then ends up being 12 times smaller than the GPT LLMs. And again, commercially, we don’t really have anyone to compare against because that isn’t available just yet. There is a saying in the Southern African context which is inspired by the philosophy of Ubuntu, which is I am because we are. And this saying is that if you want to go fast, go alone. But if you want to go far, go together. So this is an invite for you to join us in this mission in helping the future arrive well. And there are a couple of pointers that I’ve touched on throughout this presentation, but there are a couple that I would like to hone in on more specifically. The major challenges that we face in building technology for African languages and generally speaking languages in the global South that don’t have huge text bodies of evidence for multiple reasons, maybe they’re an oral heritage and an oral culture, we are facing a number of things that make our jobs very difficult, very cool and exciting, but very difficult. So for example, in terms of the data aspect, this is a stat that is taken from a Wikipedia page that summarizes the languages of which the internet consists of. There isn’t any African language on that list. I think it summarizes it to 34 or 38 languages. So we just don’t have the option to scrape the internet for information around our language. We’re having to very purposefully curate data sets, build data sets, and be very smart about how we use that data in order to train our models. But we need a lot of help there. In terms of the compute power, we often have to wait for the US to go to sleep in order to be able to use the compute powers available commercially. But on the African continent so far, there is only one commercially available data center with GPUs. And then in terms of organizations like ours, who are really trying to be here on the ground, solve the problems of the people that are within our communities, we’re not being funded very well. Only 2.2% of startups on the African continent focusing on AI and IoT are funded by VCs. So in terms of joining the mission, I’ve summarized it into these points. First and foremost, build with us, not for us. It is very difficult to build for people if they’re not at the table to help you understand what it is that needs building, how it will work, when it will work, and how you can measure its success. So make sure we’re in the room with you when you’re trying to help us in our quest for equity around language solutions. You cannot open source your way out of the problem. So democratizing access to open source models is great. But it doesn’t mean very much unless you’re also democratizing compute or supporting communities to generate their data, own their data, and be able to license their data in order to improve their NLP ecosystems. You can lay down internet cables, and that is fantastic. But what is even cooler is if the information going down those cables is something that the end consumer can consume. At the moment, you have schools all across the African continent, all across the global South, that have access to internet, but do not have access to any content in a language that they can understand. Third and foremost, support content creation, content translation, and application building. And then last off, join us in building essential smaller models. The excess use of resources has put an extensive strain on Africa, our social systems, on climate change, et cetera. It is not necessary to build all of these really large, powerful machines when we know that we can solve problems with far less. So to conclude, the African language problem is not solved. And there is no solving the SDGs without solving the language problem. And when we say here there is no solution without Africans, it’s not as a war cry. It’s like, invite us to the table, and we can work together and go far together. Thank you very much.