Active Learning Based Labeling Method for Fault Disposal Pre-plans

Power grid companies have vast quantities of data that can be mined for valuable information. However, there is a limited quantity of labeled pre-plans data due to the high labeling cost. This paper proposes an active learning-based sequence labeling method, aiming at achieving good deep learning model performance with only a small amount of labeled data. It proposes a sample selection strategy based on uncertainty and slot, in which the uncertainty and slot respectively represent the informativeness and representativeness of the sample. Meanwhile, to avoid getting into local details in each round of model training, a method for evaluating the performance of data labeling is proposed. Based on these two works, the active learning and data labeling processes are also present. We use four strategies as comparative methods in our experiment, and the results demonstrate that our method outperforms other active learning methods. We conclude that our method is efficient and applicable to tasks involving the labeling of textual data in other contexts.