Extracting landslide geological disaster relationships based on knowledge graph using deep learning approach

Considering the complexity, suddenness and spatial–temporal nature of landslide hazards, a knowledge graph provides knowledge support for geological hazards, and relationship extraction is the core of constructing the graph. However, the accumulation of unstructured data over the years obscures relationships and cross-textual connections between information, making it challenging to gain a comprehensive understanding of the hazards. To address this challenge, this study proposes a framework for extracting relationships of landslide geological hazards based on a domain ontology. First, an ontology of landslide hazard chains is created from the collected data, with concepts and relationships standardised using Protégé to produce a structured semantic representation. Next, a deep learning model is utilised for the relationship extraction task, incorporating the A-Lite Bert (ALBERT) model for character vectorisation. The textual features extracted are then input into a Bidirectional Gated Recurrent Unit-Attention (BiGRU-Attention) model for training. The resulting probabilistic weights are summed with the product of the states of the individual hidden layers to determine the result of the relationship categorisation. The experimental results show that the ALBERT-BiGRU-Attention relationship extraction model performs best, with a precision rate of 86.40%, a recall rate of 87.88%, and an F1 score of 88.46%. Therefore, the methodology outlined in this study provides technical support for the landslide hazard knowledge graph. Furthermore, the visualisation of the results highlights the spatial and temporal variability of these hazards and their impact on human activities.