Yankari: Monolingual Yoruba Dataset

Maro AkpobiACL 2025

TL;DR

This paper introduces Yankari, a large-scale, ethically sourced monolingual Yoruba dataset designed to address the critical lack of resources for this widely spoken African language in Natural Language Processing (NLP). By carefully curating over 51,000 documents from diverse, non-religious sources, Yankari aims to provide a balanced and authentic representation of contemporary Yoruba, enabling the development of more accurate and culturally appropriate NLP models.

What Problem Does This Paper Address?

Natural Language Processing (NLP) has seen remarkable advancements, but these benefits have largely been concentrated on high-resource languages. Many African languages, including Yoruba, which is spoken by over 30 million people, remain severely underrepresented in NLP research and applications. Existing Yoruba datasets often suffer from significant limitations: they are either too small, heavily skewed towards religious content (like Christian texts), or contain problematic machine-translated data, raising ethical and legal concerns. This lack of comprehensive, ethically sourced, and domain-diverse data hinders the development of effective NLP tools for Yoruba, perpetuating a digital divide.

Methods

The creation of the Yankari dataset involved a meticulous, multi-stage process prioritizing ethical considerations and data quality.

  • Source Selection: The authors carefully selected 13 diverse sources, including news outlets, blogs, educational websites, and Wikipedia, to ensure a representative sample of contemporary Yoruba.
  • Data Collection: Content was gathered from these selected sources.
  • HTML Parsing and Text Extraction: A robust system using BeautifulSoup was employed to extract clean text from web pages, preserving document structure and paragraph boundaries.
  • Deduplication: A two-step deduplication process was used: exact matching at the document level and near-duplicate detection at the paragraph level using MinHash and Locality-Sensitive Hashing (LSH).
  • Ethical Filtering: Explicit efforts were made to remove data from restricted sources, suspected machine-translated content, and inappropriate or offensive material. This filtering was largely a manual process conducted by a native Yoruba speaker to ensure linguistic authenticity and suitability for public research.
  • Quality Assurance: Native Yoruba speakers were involved in data cleaning and validation. Regular spot checks and a final manual review of a subset of the data were performed to verify quality and authenticity.

Datasets

The paper introduces the Yankari dataset, a newly created monolingual Yoruba corpus.

  • Yankari Dataset:
    • Size: 51,407 documents, totaling over 30 million tokens.
  • Sources: 13 diverse sources, including Wikipedia (32.70%), Alaroye.org (20.49%), BBC.com (16.05%), Awikonko.com.ng (10.58%), and Yoruba.von.gov.ng (4.94%).
  • Key Differentiator: Focus on non-religious, ethically sourced content, aiming for a balanced representation of contemporary Yoruba language use.

The paper also discusses existing datasets and their limitations, including:

  • Yorùbá Text C3: Noted for being heavily skewed towards religious content (Christianity) and including JW300 data, which has ethical and legal issues.
  • MENYO-20k: A multi-domain English-Yoruba corpus primarily for machine translation, but its small size limits its use for monolingual NLG tasks.
  • Wura Dataset: A multilingual dataset with approximately 68,000 Yoruba documents, but found to have critical quality issues, formatting errors, and inappropriate content, and also inherits biases from JW300 and Wikipedia.
  • Large-Scale Web-Crawled Corpora (mC4, OSCAR, Afriberta-Corpus, CulturaX): These often suffer from noise, poor formatting, limited source diversity, high duplication rates, and machine-translated content.

Key Insights and Ideas

  • The "Ethical Gap" in Low-Resource Language Data: The paper strongly emphasizes that simply collecting large amounts of data is insufficient. Ethical considerations, particularly regarding religious texts and potential cultural imperialism, are paramount for low-resource languages.
  • The Detrimental Impact of Data Repetition: Drawing on prior research, the authors highlight how data repetition can severely hinder NLP model performance by promoting memorization over generalization and obstructing the development of crucial learning mechanisms. This underscores the importance of their rigorous deduplication and quality control.
  • The Need for Non-Religious Domain Data: The paper argues that datasets heavily reliant on religious texts provide an unbalanced view of a language's usage and can be ethically problematic. Yankari's focus on diverse, non-religious domains is a key contribution.
  • Replicable Methodology for Other Languages: The authors aim for their methodology to serve as a blueprint for creating similar high-quality datasets for other underrepresented languages, promoting broader digital inclusion.

Key Results

  • The Yankari dataset comprises 51,407 documents and 30,438,702 tokens.
  • The average document length is 592.11 tokens, indicating substantial content per document.
  • The domain analysis shows a balanced distribution across encyclopedic content (Wikipedia), news (BBC, Alaroye), and cultural discussions (Awikonko, Yoruba.von.gov.ng).
  • Through rigorous analysis of existing datasets (like Wura), the authors identified critical issues such as high repetition (e.g., 18.01% of Wura containing 'asteroidi'), significant duplication (reducing Wura to 45% unique entries), and general quality issues.

Limitations and Future Work

The authors acknowledge several limitations and potential biases in the Yankari dataset:

  • Internet Bias: The dataset primarily relies on online sources, which may not fully represent all aspects of the language.
  • Written Language Bias: The dataset does not include samples of spoken Yoruba.
  • Source Bias: The selection of sources might introduce inherent biases.
  • Temporal Bias: The dataset reflects content available up to the time of collection.
  • Demographic Skew: Internet access and content creation are not uniformly distributed across all Yoruba-speaking demographics.
  • Informal Variants: Internet sources may favor more formal language use.
  • Diacritization Challenges: Tonal ambiguity and standardization issues in Yoruba orthography could lead to inconsistencies.
  • Exclusion of Non-Yoruba Characters: Content involving code-switching or borrowings might have been inadvertently removed.

Future work will focus on:

  • Rigorously evaluating Yankari's utility in downstream NLP tasks like language modeling, machine translation, and text classification.
  • Conducting comparative performance analyses against other Yoruba corpora.
  • Expanding the dataset to include more diverse content and potentially spoken language samples.

Summary updated

BibTeX
@inproceedings{akpobi2025yankari,
  author = {Maro Akpobi},
  title = {{Yankari: Monolingual Yoruba Dataset}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure