NLP for African Languages

Course outline and learning expectations

This is a tutor-led course. The outline shows what your tutor will cover; teaching materials and examinations are provided directly to enrolled students.

What you will learn

• Explain the major data, modelling and evaluation gaps in African-language NLP
• Design ethical data collection and annotation workflows for Luganda, Kiswahili and other low-resource languages
• Normalize spelling variation, code-switching, Unicode, emojis and noisy local-language text
• Evaluate mBERT, XLM-R and relevant African-focused alternatives for a defined task
• Fine-tune and assess a sentiment or text-classification model using suitable metrics and error analysis
• Identify key speech-to-text data, transcription and evaluation requirements for local languages
• Compare machine-translation tools and document bias, dialect, domain and quality limitations
• Build, document and present a reproducible local-language NLP capstone tool

Modules

1. 1

   African-language NLP landscape

   Assess the current state of NLP for African languages and identify realistic opportunities and constraints

   Language coverage and data gapsLow-resource NLP constraintsOpportunities in Uganda, East Africa and AfricaEthics, inclusion and community participation
2. 2

   Building small local-language datasets

   Plan and create small, well-documented datasets for practical local-language NLP tasks

   Use-case definitionData sourcing for Luganda and KiswahiliConsent, licensing and privacyAnnotation guidelinesQuality checks and dataset documentation
3. 3

   Text normalization for local languages

   Prepare noisy, multilingual and code-switched text for modelling without removing important linguistic meaning

   Unicode and encodingSpelling variants and dialectsTokenization and morphologyCode-switchingURLs, emojis, hashtags and noisy textReproducible preprocessing
4. 4

   Multilingual pretrained models

   Use multilingual pretrained models as baselines for African-language NLP tasks and evaluate their language coverage

   mBERT and XLM-R foundationsTokenizers and language coverageHugging Face pipelinesBaseline experimentsAfrican-focused model alternativesModel selection under compute limits
5. 5

   Fine-tuning for sentiment and classification

   Fine-tune and evaluate a multilingual model for sentiment analysis or another local-language classification task

   Label designTrain-validation-test splitsFine-tuning workflowClass imbalanceAccuracy, precision, recall and macro F1Confusion matrices and error analysis
6. 6

   Speech-to-text considerations

   Evaluate the data, linguistic and deployment issues that affect speech recognition for local languages

   Audio collection and consentTranscription conventionsAccents, dialects and code-switchingNoise and recording conditionsWord error rate and character error rateHuman review and deployment limits
7. 7

   Translation tools and limitations

   Compare translation workflows for African languages and document where automated outputs require human review

   Luganda-English and Kiswahili-English workflowsParallel data and evaluationBLEU, chrF and human evaluationBias and named entitiesDomain and dialect limitationsPost-editing and responsible use
8. 8

   Capstone: local-language NLP tool

   Build and present a small, reproducible NLP tool for an African-language use case

   Problem definitionDataset cardModel training and evaluationSimple interface or APIDocumentation and demoCapstone presentation