A multi-layer deep learning framework for joint extraction of geological disaster entities and relations

The extraction of geological disaster entities and relations is a pivotal technology in the construction of a geological disaster knowledge graph. This is significant for enhancing disaster response efficiency and fortifying disaster management. Existing methods often struggle with issues such as insufficient modeling of contextual information, error propagation, and difficulty in identifying complex relationships. To address these challenges, this paper proposes a multi-layer deep learning framework for joint extraction of geological disaster entities and relations. The framework integrates RoBERTa with a context dependency enhancement module to capture long-distance dependencies and improve word vector representation. Entity extraction adopts a bi-directional approach, predicting entity pairs based on head and tail entities to reduce error propagation. For relation extraction, the Coordinate Attention mechanism enhances entity pair features, while Biaffine's multi-relationship dynamic allocation mechanism assigns relationship types. A Chinese geological disaster entity and relationship extraction corpus (GDERECorpus) was constructed for model training and evaluation. The proposed model achieves F1 scores of 84.87% on GDERECorpus and 81.73% on DuIE, outperforming baseline models by 1.07%−2.88% and 1.06%−2.72%, respectively. These results demonstrate the model's effectiveness in extracting geological disaster-related knowledge, offering a valuable tool for disaster management and response.