AI in Africa: Hey Alexa – do you speak IGBO?1. May 2020
AI in Africa: Hey Alexa – do you speak IGBO?
London, April 1, 2020
Bonaventure Dossou has thought a lot about how to improve the telephone conversations with his mother. She often sends him voice messages in Fon, a Beninese language, since he is currently studying in Russia. However, he does not understand some of the expressions she uses. “My mother cannot write Fon and I don’t speak the language very well, but I am fluent in French,” Mr Dossou told the BBC. “I often ask my sister to help me understand some of the expressions my mother uses,” he said.
Fon sentences in English
Nùkócé nɔn yìnMy name is
Oun yìn wàn nouwé I love you
Ouh fɔn gangjiI is doing well
Source: Bonaventura Dossou
An improvement of his Fon through study is out of the question, because like hundreds of other African languages it is mostly spoken and rarely documented, so there are few, if any, books for teaching grammar and syntax.
Driven by curiosity and driven by data integrated from a Fon into a French Jehovah’s Witnesses’ Bible, Mr. Dossou and Chris Emezue, a Nigerian friend, developed an artificial intelligence (AI) language translation model similar to Google Translate, which they called FFR. It is still in progress.
The two students are among several AI researchers who use African languages in Natural Language Processing (NLP), a branch of AI designed to teach and assist computers to understand human languages.
If the world had not come to a standstill after the Covid 19 pandemic, Mr. Dossou and Mr. Emezue would have presented their creation to hundreds of participants at one of the world’s largest AI conferences, the ICLR, in Ethiopia’s capital Addis Ababa this week.
It would have been the first time that the event had taken place in Africa.
Instead of canceling the event, the organizers decided to hold it virtually.
AI innovations were highlighted as the driving force behind the so-called fourth industrial revolution, which will radically change almost every aspect of our lives, including the way we work.
At present, Africa is seen as the loser in shaping the future of AI, as the majority of the continent’s estimated 2,000 languages are classified as “resource poor”, meaning that there is a lack of data on them and/or that what is available has not been indexed and stored in formats that can be useful.
This was a severe blow
African languages are not considered in the creation of NLP applications such as language assistants, image recognition software, traffic warning systems and others.
But African researchers are working to overcome this handicap.
“We are focusing on putting Africa on the map of NLP and AI research,” Dr Ignatius Ezeani of the University of Lancaster told the BBC.
“Until you have your language resources publicly, freely and openly available, researchers will not have the data for creative solutions in the blink of an eye. We will always have to rely on, say, Google to determine the direction of research,” said Dr Ezeani.
The conference in Ethiopia should be a big deal for African researchers who, among other things, have been denied visas to attend previous ICLR conferences in the US and Canada, excluding them from global AI talks.
“Not having the conference in Addis was a severe blow, it would have meant a massive change in the diversity of the conference,” Jade Abbott, founder of Masakhane, a research movement for machine translation for African languages, told the BBC.
Most of the time we start from scratch
Masakhane, which in isiZulu means “We build together”, has 150 members in 20 African countries. Membership is open to anyone interested in language translation.
“We are building a community of people who care about African languages and are interested in building translation models. 30% of the world’s languages are African, so why don’t we have 30% of NLP publications? asked Ms Abbott.
The network focuses on promoting language translation from Africans for Africans by Africans and encourages open sharing of resources and collaboration so that researchers can build on each other’s work.
Most of the time, however, it means starting from scratch. For example, a researcher associated with Masakhane is currently collecting data from speakers of the Damara, a Khoisan language – famous for its clicking sound – in Namibia, Ms Abbott said. So far, Masakhane affiliates have produced 35 translations into 25 African languages, she added.
In addition to Masakhane, there are other initiatives to build and strengthen networks of AI researchers on the continent:
– Deep Learning Indaba, which promotes AI in Africa and holds an annual conference
– Data Science Africa, which connects the continent’s researchers
– BlackinAI, an initiative that promotes the involvement of black people in the field of artificial intelligence
Dr Ezeani calls them “silent battles” of Africans working in the field of AI.
He sees these engagements as a contribution to the expansion of the continent’s capacities, both in terms of building the AI infrastructure and the skills of researchers and developers. “This is essential not only for recognition, but also for actually addressing our local challenges, e.g. in health, agriculture, education and governance, with home-grown and targeted solutions,” he said. “Perhaps at some point we can also take responsibility and control the narrative,” he added.
Hey Alexa, do you speak Igbo?
Dr. Ezeani is currently working on a machine translation of the Nigerian Igbo language into English. “In five to ten years, I think I will be able to interact with Alexa in Igbo or any minority language, which will be a great and fulfilling achievement,” said Dr. Ezeani.
Currently, neither Alexa from Amazon, nor Siri from Apple, nor Google Home, the major players in the global market for language assistants, support a single African native language. Google Translate is enabled for 13 African languages, including Igbo, but it is far from perfect.
Dr Ezeani said that the work he and others are doing could tempt technology companies to integrate African languages into their devices. However, he warns that African researchers working in the field of AI should be guided by original ideas “that are really useful to people” and not pursue vanity projects. “We can check whether, for example, a translation from Igbo to Yoruba and vice versa is actually more useful than a translation from Igbo into English, or whether linguistic or visual systems are more necessary than text-to-text,” he said.
As for Mr Dossou and his co-creator, Mr Emezue, they have great ambitions for the FFR if they can secure funding. They see Fon, a Bantu language spoken by more than two million people in Benin, but also in parts of Nigeria and Togo, as helping them to extend their work to other markets.
Fon is part of the Niger-Congo language family, which means it shares a common lineage with languages spoken in parts of western, central, eastern and southern Africa. For the time being, however, their main focus is to continue training FFR to better translate daily conversation. “Perhaps next year or so, my mother’s messages in Fon will be translated into French texts,” said Mr. Dossou.