A Pipeline Approach for Entity and Relationship Extraction in Geological Hazard Texts Using BERT-BiLSTM-CRF with Self-Attention

Geohazard knowledge is a vital component of geological knowledge bases, playing a crucial role in disaster prevention and risk assessment. However, existing information extraction methods struggle with long-distance dependencies in geohazard texts, leading to poor performance in extracting relevant entities and their semantic relationships. To address this, we propose a pipeline approach based on deep learning models for enhanced geohazard information extraction. First, the BERT-BiLSTM-CRF model is applied to extract five types of entities in the geohazard domain, achieving an average F1-score of 90.22%. This demonstrates the model’s ability to effectively capture the semantic meaning of entities in geohazard texts. Next, to tackle the issue of long-distance semantic dependencies, a self-attention mechanism is introduced in the relation extraction task, combined with the BERT-BiLSTM-CRF model. The F1-score for relation extraction reaches an average of 82.7%, significantly improving the model's stability in recognizing relationships in long texts. Experimental results show that the proposed method outperforms traditional approaches in both entity and relation extraction, thus improving the accuracy of geohazard knowledge extraction. These findings promote the intelligent application of geohazard information and further enrich and refine geological knowledge bases.