Automatic Extraction of Flooding Control Knowledge from Rich Literature Texts Using Deep Learning

Flood control is a global problem; increasing number of flooding disasters occur annually induced by climate change and extreme weather events. studies are important knowledge sources for flood risk reduction have been recorded in the academic literature. The main objective this paper was to acquire from long-tail data literature using deep learning techniques. Screening conducted obtain 4742 flood-related documents past two decades. Machine parse documents, 347 sample points different years were collected sentence segmentation (approximately 61,000 sentences) manual annotation. Traditional machine (NB, LR, SVM, RF) artificial neural network-based algorithms (Bert, Bert-CNN, Bert-RNN, ERNIE) implemented model training, complete sentence-level extraction batches. results revealed that methods exhibit better performance than traditional terms accuracy, but their training time much longer. Based on comprehensive feature capability computational efficiency, performances ranked as: ERNIE > Bert-CNN Bert Bert-RNN. When as benchmark model, several deformation models showed applicable characteristics. Bert, Bert-RNN good at acquiring features, local processing variable-length inputs, respectively. improved masking mechanism corpus therefore exhibited performance. Finally, 124,196 usage method 8935 quotation sentences obtained proportions trends over last 20 years. Thus, with more accumulates, study lays foundation future.