Cold-start Active Learning through Self-supervised Language Modeling
This work addresses the challenge of reducing annotation costs in NLP for practitioners, though it is incremental as it adapts existing pre-trained models to a specific setting.
The paper tackles the cold-start problem in active learning for text classification by using pre-trained language model loss as a proxy for uncertainty, achieving higher accuracy with fewer labeling iterations and less computation time compared to baselines.
Active learning strives to reduce annotation costs by choosing the most critical examples to label. Typically, the active learning strategy is contingent on the classification model. For instance, uncertainty sampling depends on poorly calibrated model confidence scores. In the cold-start setting, active learning is impractical because of model instability and data scarcity. Fortunately, modern NLP provides an additional source of information: pre-trained language models. The pre-training loss can find examples that surprise the model and should be labeled for efficient fine-tuning. Therefore, we treat the language modeling loss as a proxy for classification uncertainty. With BERT, we develop a simple strategy based on the masked language modeling loss that minimizes labeling costs for text classification. Compared to other baselines, our approach reaches higher accuracy within less sampling iterations and computation time.