🧑‍🔬 Jesujoba Alabi

Jesujoba Alabi works in natural language processing for languages that large models serve poorly, with a strong focus on African languages. Their papers here ask how pre-trained multilingual models can be extended to languages they barely saw in training, how scarce annotated data can be stretched through cross-lingual transfer, and how to build test sets that let researchers measure progress on many languages at once. The methods combine continued pre-training, fine-tuning on small curated datasets, and the construction of human-annotated corpora and benchmarks, placing the work within Multilingual & cross-lingual NLP.

The most cited paper, Adapting Pre-trained Language Models to African Languages via Multilingual Adaptive Fine-Tuning, replaces one-model-per-language adaptation with a single model trained jointly on many African languages alongside a few high-resource ones. It also trims vocabulary for unneeded scripts to shrink the model considerably. A Few Thousand Translations Go a Long Way! examines how pre-trained translation systems can be extended to unseen African languages and shifted from mostly religious parallel text toward the news domain. MasakhaNER 2.0 provides a news-domain named entity corpus for languages across Sub-Saharan Africa and studies which source languages transfer best, testing linguistic and geographic predictors.

The newest work moves from individual tasks toward broad evaluation. SIB-200 builds a topic classification benchmark covering more than two hundred languages and dialects by labelling English sentences from a multi-way parallel corpus and projecting the labels onto every translation. This gives many low-resource languages their first public test data for language understanding.

Alabi is based at Saarland University, with earlier bylines at the Max Planck Institute for Informatics, the German Research Centre for Artificial Intelligence and the University of Ibadan. The most frequent collaborator on these papers is David Ifeoluwa Adelani. Other co-authors include Dietrich Klakow and Peter Nabende.

Jesujoba Alabi has 4 papers in our library, in 2022–2023, most of them on Multilingual & cross-lingual NLP, Named entity recognition, Language models & pretraining; often with David Ifeoluwa Adelani, Dietrich Klakow, Peter Nabende.