How Much Swahili Data Is Enough? Data-Efficient English-to-Swahili Transfer for Inclusive Intent Classification

Intent classification is an important component of digital assistants and service-routing systems, yet building reliable models for Swahili remains difficult when task-labeled data are limited while equivalent English data are more readily available. This challenge is relevant to language access across East Africa, including Muslim communities that use Swahili in everyday communication and increasingly interact with digital services. 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. Rather than assuming that larger multilingual models are always preferable, we examine how model choice changes with the availability of target-language labels and computational resources. The study provides an empirical basis for developing more accessible Swahili language technologies when local annotation capacity is limited.