AfriSUD: New Treebank Exposes AI's Syntax Gap on 9 African Languages

A team of 18 researchers led by Happy Buzaaba has released AfriSUD, the first large-scale collection of syntactically annotated treebanks covering nine diverse African languages. Spanning major language families and regions across Sub-Saharan Africa, the dataset uses the Surface-Syntactic Universal Dependencies (SUD) framework and is verified by native speakers to ensure high quality. It captures typological features such as agglutination and tone, which are often underrepresented in NLP resources.

Evaluating models on part-of-speech tagging and dependency parsing, the team tested non-transformer baselines, multilingual pretrained encoders (like mBERT), and LLMs. Results show a clear syntax gap: even state-of-the-art models fail to fully handle the structural diversity of African-language syntax. This work highlights a critical blind spot in AI—most benchmarks focus on high-resource languages, leaving African languages underserved. AfriSUD provides a robust tool for measuring and improving model performance on these languages.