Language Justice in the Age of AI

Artificial intelligence has made remarkable strides in natural language understanding over the past decade. Systems like GPT-4, Gemini, and Claude demonstrate near-human performance on English text — but this performance rests on an enormous disparity: English, Chinese, French, Spanish, and a handful of European languages account for the overwhelming majority of training data on which today's foundation models are built.

Africa presents the most acute expression of this gap. The continent's estimated 2,000 to 3,000 languages include some of the world's most complex linguistic structures, yet fewer than 20 — and by some measures, fewer than ten — receive substantive representation in published NLP research or commercial AI products. Hausa, Swahili, Yoruba, Amharic, and Zulu represent the best-resourced end of a very sparse spectrum. Languages like Igbo, Twi, Dholuo, Wolof, Ndebele, and hundreds of others remain effectively invisible to AI systems.

This is not merely a technical inconvenience. It is a structural injustice. When citizens cannot access government services, legal information, healthcare guidance, or educational materials in their own language, they are excluded from the digital economy regardless of device ownership or internet access. When language AI is built without African representation, the systems that shape daily life — content moderation algorithms, loan eligibility models, medical diagnostic tools — fail disproportionately for African users.

This paper documents TNG's experience building a Yoruba-language legal translation system for Lagos State Government and derives from that experience a generalised framework applicable across African language contexts.

To build NLP systems for African languages, developers must first understand why standard approaches underperform. Three structural characteristics create the primary challenges: tonal phonology, agglutinative morphology, and data scarcity.

Tonal languages, including Yoruba, Igbo, Ewe, and many others, assign meaning to syllables partly through pitch. The Yoruba word 'owo' means hand, money, or respect depending on tonal marking. Standard tokenisers built for non-tonal languages strip or mishandle diacritical marks, creating semantic ambiguity that propagates through downstream tasks. A model that cannot distinguish tonal variants cannot reliably translate or classify text in these languages.

Agglutinative morphology, prominent in Swahili, Zulu, Xhosa, and Bantu languages broadly, allows a single word to encode what English would express in an entire phrase. 'Nitakupendeza' in Swahili means 'I will please you' — one token that standard subword tokenisers built on English vocabulary will fragment poorly. The result is vocabularies with artificially high out-of-vocabulary rates and models that must process African language text through a lens designed for morphologically simpler languages.

Data scarcity is the most immediate and practically solvable of the three. As of 2025, high-quality annotated corpora for most African languages number in the tens of thousands of sentences rather than the billions available for English. This is partly a historical artefact — colonial language policy systematically marginalised indigenous writing systems — and partly a consequence of continued underinvestment.

In 2024, TNG was commissioned by Lagos State Government to build an AI-powered translation pipeline for the Lagos State Laws — a corpus of 70+ statutes covering commercial law, land use, environmental regulation, traffic enforcement, and public health.

The project faced three simultaneous challenges: the source documents were in dense legal English; the target language (Yoruba) had limited NLP tooling; and the accuracy requirements for legal text are exceptionally high — a mistranslation of a statute is not merely inaccurate, it is potentially harmful to citizens who rely on that text to understand their rights and obligations.

Our pipeline combined fine-tuned multilingual transformer models (specifically mT5 and NLLB-200, both adapted on Yoruba legal corpora we assembled) with a structured human review process involving Yoruba-speaking legal professionals. Automated translation was used for first-pass output; human annotators corrected and refined each passage; corrected outputs were fed back into model fine-tuning in an iterative loop.

We achieved 95% translation fidelity as measured by bilingual evaluation understudy (BLEU) scores against expert human reference translations, with near-perfect recall on key legal terms across 15 domains of law. The full pipeline processed the 70+-statute corpus in under four months — a timeline that would have been impossible with purely manual translation.

From the Lagos project and subsequent work, we propose a four-pillar framework for building reliable NLP systems for low-resource African languages.

Pillar One — Corpus Development as Infrastructure: No model can outperform its training data. For African languages, investing in corpora is not a project cost — it is infrastructure investment equivalent to building a road. Corpora must be domain-specific (general text corpora do not produce legal-grade translation accuracy), community-annotated (native speakers must be paid appropriately for annotation work), and openly licensed so that the investment compounds across subsequent projects.

Pillar Two — Transfer Learning from Linguistic Relatives: Full training from scratch is prohibitively expensive for most African language NLP projects. Transfer learning from pre-trained multilingual models (mBERT, XLM-R, NLLB-200) provides a viable starting point, particularly when source and target languages share significant structural features. Yoruba benefits from Bantu-adjacent transfer; Hausa from Arabic-script adjacency.

Pillar Three — Native Speaker Expertise at the Centre: AI-generated language output must be evaluated by native speakers, not by automated metrics alone. This is both a quality imperative and an ethical requirement — the communities whose languages are being processed must have meaningful authority over the outputs. This means hiring annotators appropriately, building feedback mechanisms into deployed systems, and treating community expertise as a valued input rather than a validation step.

Pillar Four — Governance Before Deployment: Language AI deployed in high-stakes contexts (law, healthcare, education) must have clear accountability structures. Who owns the model? Who can audit its outputs? What happens when it is wrong? In public sector deployments, these questions must be answered in policy before deployment, not retroactively.

Technical frameworks alone are insufficient. The language gap in AI will not close without policy intervention at national and continental levels.

African governments should designate open language corpora as digital public goods and fund their development accordingly. The cost of a well-annotated, publicly licensed corpus for a major African language is a fraction of the cost of the government services that would benefit from that corpus — yet no African government currently funds corpus development at scale.

The African Union's AI strategy, published in 2024, identifies language inclusion as a priority but does not specify funding mechanisms or timelines. We call for AU member states to establish a Continental Language AI Fund, modelled on existing digital public goods initiatives, to co-fund corpus development and share resulting datasets across borders.

Technology companies operating in African markets should be required, as a condition of market access, to publish language support roadmaps that include specific, time-bound commitments to African language inclusion. Voluntary pledges have not moved the needle.

Universities across Africa should establish Language Technology research centres connected to government and civil society users — ensuring that research produces deployable systems rather than papers that do not leave the academy.