Multi-task Active Learning for Pre-trained Transformer-based Models

Abstract Multi-task learning, in which several tasks are jointly learned by a single model, allows NLP models to share information from multiple annotations and may facilitate better predictions when the inter-related. This technique, however, requires annotating same text with annotation schemes, be costly laborious. Active learning (AL) has been demonstrated optimize processes iteratively selecting unlabeled examples whose is most valuable for model. Yet, multi-task active (MT-AL) not applied state-of-the-art pre-trained Transformer-based models. paper aims close this gap. We explore various selection criteria three realistic scenarios, reflecting different relations between participating tasks, demonstrate effectiveness of compared single-task selection. Our results suggest that MT-AL can effectively used order minimize efforts models.1