Scaling African Language Datasets

Recent efforts to create more language resources for Africa have been spearheaded by local communities. While this has improved representation for major African languages, significant gaps persist for the 2,000+ African languages that remain severely low-resource. These initiatives have also highlighted a critical issue – the sustainability and continuity of research efforts are often short-lived. To ensure the longevity of African datasets, encouraging open contributions is essential. Open collaboration allows datasets to evolve far beyond their original scope and incorporate the natural evolution of language over time.

Dataset curation often faces significant challenges due to the high costs and scalability issues associated with manual human effort. Community-driven efforts produce high-quality, realistic data but suffer from scalability challenges, including cost and uneven contribution distribution. Conversely, synthetic data generation using multilingual models offers greater scale but introduces validity issues, such as artifacts that diverge from native phrasing, limited real-world complexity representation, and the risk of bias amplification from web-trained models that underrepresent African languages. Therefore, while synthetic data generation is more scalable, it must be complemented with human quality checks to ensure the inclusion of cultural expression and fair representation.

At the core of Tonative lies the aim to reduce manual human labour while maintaining a crucial human-in-the-loop approach. It extends existing multilingual datasets into more African languages through three interconnected stages:

I Identifying high-quality data resources for extension; II Leveraging multilingual models for initial translations; III Validating translations with native speakers and linguists.

For instance, we extended Mbogho et al.’s dataset of 29,280 sentences, originally comprising Kidaw’ida-Kiswahili, Kalenjin-Kiswahili, and Dholuo-Kiswahili corpora, by adding a Kiswahili-English component. This extension facilitates cross-lingual transfer and lays the groundwork for future translations into other languages. With volunteer data validators already in place, we ensure the quality and authenticity of translations. We adopted a similar methodology with Masakhane’s AfriXNLI, an African adaptation of Facebook’s XNLI dataset, which initially consists of 7,500 samples. Our contribution involved significantly expanding this dataset. For some of the languages already present in AfriXNLI, we increased the sample size from the original 1,050 samples per language to 5,000 samples. Furthermore, we extended the dataset to include new African languages that were not part of the initial AfriXNLI effort.

While we shall continue to explore more optimised methods of data curation, our current approach has demonstrated a significant improvement in dataset extension for African languages through human-AI collaboration. By extending existing datasets, leveraging pretrained multilingual models, and embedding community validation, we provide a more sustainable pathway to increasing data resources in Africa.

Resources and Links