Exploring the Efficiency of Batch Active Learning for Human-in-the-Loop Relation Extraction

Domain-specific relation extraction requires training data for supervised learning models, and thus, significant labeling effort. Distant supervision is often leveraged for creating large annotated corpora however these methods require handling the inherent noise. On the other hand, active learning approaches can reduce the annotation cost by selecting the most beneficial examples to label in order to learn a good model. The choice of examples can be performed sequentially, i.e. select one example in each iteration, or in batches, i.e. select a set of examples in each iteration. The optimization of the batch size is a practical problem faced in every real-world application of active learning, however it is often treated as a parameter decided in advance. In this work, we study the trade-off between model performance, the number of requested labels in a batch and the time spent in each round for real-time, domain specific relation extraction. Our results show that the use of an appropriate batch size produces competitive performance, even compared to a fully sequential strategy, while reducing the training time dramatically.