LHAKG: a knowledge graph construction framework for landslide hazard assessment by using XLNet-BiLSTM-CRF from geoscience literature

Reliable landslide hazard assessment in complex environments is essential for effective risk prevention and control. However, extracting and integrating relevant knowledge from extensive, unstructured geoscience literature remains a significant challenge. This paper proposes a framework to construct a landslide hazard assessment knowledge graph (LHAKG). First, the framework develops an ontology comprising four core elements: spatial information, causative factor, assessment model, and assessment data. These elements and their interrelations form the conceptual backbone of the LHAKG. Next, an XLNet-BiLSTM-CRF model is trained on a self-developed domain-specific corpus to extract entity information related to landslide hazard assessment elements from scientific literature. Finally, Sentence-Transformer techniques are applied for knowledge fusion, facilitating semantic alignment and refinement of entities across heterogeneous categories. A case study using a dataset of 1,612 geoscience papers demonstrates the effectiveness of the framework, yielding the LHAKG with 13,400 nodes and 62,975 relationships. The results confirm that LHAKG can efficiently extract, organize, and represent landslide hazard assessment knowledge. Furthermore, the LHAKG facilitates multi-dimensional knowledge recommendation across diverse methodologies and geographic contexts. This capability enhances intelligent analysis and decision support in geohazard research, especially with incomplete data, establishing a robust foundation for broad-scale landslide risk assessment and prediction.