Revisiting Uncertainty-based Query Strategies for Active Learning with Transformers

Findings of the Association for Computational Linguistics: ACL 2022, pages 2194 - 2203 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Revisiting Uncertainty-based Query Strategies for Active Learning with Transformers Christopher Schröder, Andreas Niekler, and Martin Potthast Leipzig University Abstract Active learning is the iterative construction of a classification model through targeted la- beling, enabling significant labeling cost sav- ings. As most research on active learning has been carried out before transformer-based lan- guage models (“transformers”) became popu- lar, despite its practical importance, compara- bly few papers have investigated how trans- formers can be combined with active learning to date. This can be attributed to the fact that using state-of-the-art query strategies for trans- formers induces a prohibitive runtime over- head, which effectively nullifies, or even out- weighs the desired cost savings. For this rea- son, we revisit uncertainty-based query strate- gies, which had been largely outperformed be- fore, but are particularly suited in the context of fine-tuning transformers. In an extensive evaluation, we connect transformers to exper- iments from previous research, assessing their performance on five widely used text classifi- cation benchmarks. For active learning with transformers, several other uncertainty-based approaches outperform the well-known predic- tion entropy query strategy, thereby challeng- ing its status as most popular uncertainty base- line in active learning for text classification. 1 Introduction Collecting labeled data for machine learning can be costly and time-consuming. A key technique to minimize labeling costs has been active learning, where an oracle (e.g., a human expert) is queried to label problem instances selected that are deemed to be most informative to the learning algorithm’s next iteration according to a query strategy. Active learning is characterized by the real- world machine learning scenario in which large amounts of training data are unavailable, which may explain why comparably little research has investigated deep learning in this context. The re- cent widely successful transformer-based language models can circumvent the limitations imposed by small training datasets (Vaswani et al., 2017; De- vlin et al., 2019). Pre-trained on large amounts of unlabeled text, they can be fine-tuned to a given task using far less training data than when trained from scratch. However, their high number of model parameters renders them computationally highly expensive, for query strategies that are targeted at neural networks or text classification (Settles et al., 2007; Zhang et al., 2017), resulting in prohibitive turnaround times between labeling steps. In this paper, we systematically investigate uncertainty-based query strategies as a computa- tionally inexpensive alternative. Despite their rela- tive disadvantages in traditional active learning, when paired with transformers, they are highly effective as well as efficient. Our extensive ex- periments assess a multitude of combinations in- cluding state-of-the-art transformer models BERT (Devlin et al., 2019) and DistilRoBERTa (Sanh et al., 2019), five well-known sentence classifica- tion benchmarks, and five query strategies.1 2 Related Work Uncertainty-based query strategies used to be the most common choice in active learning, using uncertainty scores obtained from the learning al- gorithm (Lewis and Gale, 1994), estimates ob- …

Content truncated to 500 words