AfriScience-MT: Bridging the Language Gap in African Scientific Communication

Science has long been a field dominated by a handful of colonial languages, creating barriers for non-speakers. In Africa, this linguistic divide keeps millions from accessing scientific knowledge. Enter AfriScience-MT, a project aiming to bridge this gap by developing a parallel corpus for six African languages across various scientific fields.

Translation Challenges and Breakthroughs

AfriScience-MT covers Amharic, Hausa, Luganda, Northern Sotho, Yorùbá, and isiZulu, translating plain-language summaries of scientific papers into each language. This isn't just about swapping words. It's about generating new scientific terminology where none existed. An ambitious task, but one that's essential for making science truly global.

Machine translation systems and large language models were put to the test in this project. Results show a clear winner: closed-source models outperformed open-source ones. GPT-5.4 and Gemini-3.1-Flash-Lite lead the pack with sentence-level COMET scores of 68.3 and 68.0, respectively. At the document level, they tie at 48.3. Curious about the open-source performance? NLLB-1.3B, when fine-tuned, hit 67.3 at the sentence level, whereas TranslateGemma-12B managed 44.0 with 1-shot in-context learning.

Why This Matters

These numbers tell us a lot about the state of machine translation technology. The fact that closed-source models outshine open ones isn't just a fluke. It's a wake-up call about the resources and refinement these proprietary systems receive. But let's not ignore what AfriScience-MT represents, an essential move towards linguistic inclusivity in science.

Here's a pointed question: Can we claim scientific advancement when a massive portion of the global population can't even understand the conversation? AfriScience-MT is a step towards answering that, providing a scaffold for future work in African languages.

The intersection is real. Ninety percent of the projects aren't. AfriScience-MT might just be the exception, offering a genuine contribution rather than vaporware. Of course, there's the inevitable question of inference costs and who shoulders them. Translating science isn't cheap, nor is it simple, but ignoring the linguistic diversity of Africa isn't an option if we aim for inclusive growth in scientific fields.

Looking Forward

As AfriScience-MT gains traction, it's important that the open-source community takes heed. If they want to compete, they'll need to catch up quickly. The gap between closed and open-source models in this context isn't just technical. It's a gap in access and opportunity. If the AI can hold a wallet, who writes the risk model? It may not be a question of if, but when, efforts like this will reshape the scientific landscape in Africa.