🛰️ DSFSI Dispatch | May the 4th Edition
“In a research lab not so far away...”
Happy May the 4th! While Jedi Masters battled Sith Lords in the stars, our DSFSI researchers were hard at work confronting the dark side of data scarcity, language bias, and limited compute across the African continent. Today, we celebrate the Force of responsible AI through five powerful papers from the DSFSI alliance.
📡 1. Political Sentiment in the Twitter System (JeDEM Journal)
The Election Strikes Back
Penelope Matloga, Vukosi Marivate, and Kayode Olaleye explored political sentiment on X (Twitter) during South Africa’s 2021 local government elections. Using RoBERTa, VADER, and GPT-3.5, they showed how people expressed both admiration and frustration toward the ANC—and how bots tried (and failed) to sway the system.
📖 Read: https://www.jedem.org/index.php/jedem/article/view/945
🎧 Listen: https://notebooklm.google.com/notebook/f0adc2b5-b68d-4ccc-b67d-528cd62ccfd7/audio
🌐 2. Translation with the Force of Prompts (Machine Learning with Applications)
The Prompt Awakens
Pitso Khoboko, Vukosi Marivate, and Joseph Sefara trained Mistral 7B to translate English into isiZulu and isiXhosa. Using clever prompt design and QLoRA, they beat Google Translate and neared NLLB—all with minimal GPUs. The Rebellion proves once again: size isn’t everything, structure is.
📖 Read: https://doi.org/10.1016/j.mlwa.2025.100649
🎧 Listen: https://notebooklm.google.com/notebook/d9d8c38f-25b4-4dc2-bdea-9ae183e00604/audio
🌾 3. QA for Farmers on the Outer Rim (Applied AI Letters)
The Agrarian Jedi Council
Fiskani Banda, Vukosi Marivate, and Joyce Nakatumba-Nabende built a question-answering system in isiZulu, isiXhosa, Afrikaans, and English—trained on Pula Imvula farming articles. Using few-shot learning and GPT-3.5, they bring the Force of knowledge to smallholder farmers in local languages.
📖 Read: https://doi.org/10.1002/ail2.122
🎧 Listen: https://notebooklm.google.com/notebook/9ecb6a74-66a0-43fe-b7ea-bf063d9ad6e2/audio
⚖️ 4. Summarizing the Jedi Archives: ZASCA-Sum (Data in Brief)
The Republic’s Legal Records, Now Machine-Readable
Idris Abdulmumin and Vukosi Marivate present ZASCA-Sum, a landmark dataset of 4,000+ South African Supreme Court of Appeal judgments and media summaries—manually aligned and processed for legal summarization research. A resource to train legal LLMs, adapt models to South African English, and democratize access to justice.
📖 Read: https://doi.org/10.1016/j.dib.2025.111567
📂 Data: https://huggingface.co/datasets/dsfsi/zasca-sum
💫 5. A Galactic Alliance of Acknowledgements
These victories were made possible by the unwavering support of the:
ABSA Chair of Data Science
OpenAI (research compute credits)
SADiLaR, Grain SA
FCDO and IDRC (AI4D)
Legal guidance from Dr. Chijioke Okorie (Data Science Law Lab)
⚔️ This is the way.
As we build a future where African languages and legal systems are well represented in the digital world, we salute the researchers carrying the light of innovation.
May the Data be with you.
— The DSFSI Alliance