Extracting Behavior Events from Epidemic Daily Reports by Combining Deep Learning and Adversarial Training

Extracting emergencies from network information has become an important means of public security. However, existing event extraction methods rely on large-scale annotated corpora and are difficult to be effectively applied to extract epidemic behavior events, which is a brand-new task of event extraction. Firstly, an epidemic provenance model is defined to comprehensively model epidemic information in epidemic daily reports. On this basis, we propose an epidemic extraction model of epidemic behavior events, called CBEventMine, that combines deep learning and adversarial training. A relatively simple model structure of BiLSTM+CRF+sigmoid is adopted to achieve fast convergence of the model on small sample data, and an adversarial learning mechanism is added to alleviate the overfitting problem of the model in few-slot learning. The experimental results demonstrate that the proposed model CBEventMine could effectively extract epidemic event than the existing state-of-the-art methods.