West African Multilingual Annotated Text Dataset

This is a synthetic dataset generated from high-quality expert-labelled seed data. All records are algorithmically derived — statistical distributions, inter-field correlations, and annotation characteristics faithfully replicate real-world patterns from the source data, while ensuring no real individual, organisation, or transaction can be identified or reconstructed.

The West African Multilingual Annotated Text Dataset is a 5M+ token corpus covering five languages and language varieties spoken across Nigeria and Ghana: Hausa, Yoruba, Igbo, Twi, and Nigerian Pidgin. Text was collected from social media (Twitter/X, Facebook), news portals, radio transcripts, and SMS-style customer service conversations — sources that reflect the actual linguistic register of West African digital communication rather than formal written text.

Each text unit (sentence or short paragraph) is annotated with: a primary language identification label, a sentiment class (positive, negative, neutral, mixed), a named-entity recognition (NER) span set covering persons, locations, organisations, and product names, and — for conversational data — an intent label from a 24-class taxonomy aligned with common customer-service and civic-information use cases. All annotations were produced by native-speaker annotators with a three-reviewer consensus protocol; inter-annotator agreement (Fleiss κ) is reported per language and annotation type.

The dataset is structured as JSONL with one record per annotated text unit, making it compatible with standard NLP fine-tuning pipelines (HuggingFace Datasets, spaCy, OpenAI fine-tune format). Language-stratified train / validation / test splits are provided. A separate romanisation-normalised variant is available for models that struggle with code-switching and diacritic-heavy orthography.