Скачать или смотреть Active Learning for Efficient Low-Resource Efficient NLP

This presentation examines active learning as a principled framework for building efficient, cost-aware NLP systems in low-resource settings, with a particular emphasis on African languages. It demonstrates how modeling uncertainty and annotation cost enables substantial reductions in data, compute, and evaluation budgets while preserving or improving model performance. The talk is structured around three complementary lines of work. First, I present uncertainty-driven active learning for African-accented speech recognition, demonstrating how epistemic uncertainty can guide data selection to achieve significant reductions in word error rate with far fewer labeled utterances. Second, I introduce a self-active learning framework for multilingual language modeling, illustrating that models trained on orders of magnitude less data can remain competitive with large pretrained baselines across named entity recognition, text classification, and sentiment analysis. Third, I discuss cost-aware active learning via KnapsackBALD, which explicitly accounts for non-uniform annotation costs and formulates data acquisition as a budget-constrained optimization problem. The presentation further explores continual fine-tuning as a scalable alternative to retraining from scratch in active learning loops. It outlines an active learning for evaluation paradigm that adaptively selects evaluation items to reduce uncertainty in model performance estimates rapidly. Overall, the talk positions active learning as a unifying methodology for data-efficient, budget-constrained, and sustainable NLP, offering practical insights for deploying models in real-world low-resource environments where annotation effort, expertise, and computation are limited.