Semi-supervised geological disasters named entity recognition using few labeled data

The geological disasters Named Entity Recognition (NER) method aims to recognize entities reflecting disaster event information in unstructured texts to construct a geohazard knowledge graph that can provide a reference for disaster emergency response. Without training on large-scale labeled data, current NER methods based on deep learning models cannot identify specific geological disaster entities from geological disaster situation reports. However, manually labeling geohazard situation reports is tedious and time-consuming. As a result, we present Semi-GDNER, a semi-supervised geological disasters NER approach that can effectively extract six kinds of geological disaster entities when a few manually labeled and unlabeled in-domain data are available. It is divided into two stages: (1) transferring the parameters of the pre-trained BERT-base model to the BERT layer of the backbone model BERT-BiLSTM-CRF and training the backbone model with a few labeled data; (2) continuing training the backbone model by expanding the training set with unlabeled data using a self-training (ST) strategy. To reduce noise in the second stage, we select the pseudo-labeled samples with high confidence to join the training set in each ST iteration. Experiments on our constructed Geological Disaster NER data show that our approach achieves a higher F1 (0.88) than other NER approaches (including five supervised NER approaches and a semi-supervised NER approach using the ST strategy of expanding the training set with all pseudo-labeled data), demonstrating the effectiveness of our approach. Furthermore, experiments on four general Chinese NER datasets show that the framework of our approach is transferable.