How Much Swahili Data Is Enough? English-to-Swahili Transfer for Low-Resource Intent Classification

Intent classification routes a user request to the action or service that should handle it, but building this component for Swahili is difficult when task-labeled data are scarce while equivalent English data already exist. We study data-efficient English-to-Swahili transfer on MASSIVE using a lexical SVM, a BiLSTM, and XLM-R. The evaluation covers full supervision, English-only zero-shot transfer, sequential adaptation at five Swahili annotation budgets, and a controlled synthetic mixed-language stress test with simple augmentation. The goal is to identify when multilingual transfer is worth its annotation and compute cost and when a lightweight model is preferable. This resource-regime perspective is motivated by African digital services that need local-language intent routing under limited data and infrastructure.