Joint Extraction Method for Spatial Relations in Chinese Geological Texts

Abstract

Extracting spatial relations of geological entities is an important prerequisite for achieving natural language processing tasks such as geological knowledge question answering and semantic search, and is an important means to achieve structural reconstruction of unstructured geological data. Geological survey reports are standardized records and representations of geological work results. Geologists have accumulated a large amount of results data in their long-term work practice, which contains rich geological entity attribute information and spatial information. This paper uses regional geological survey reports as data sources to construct a Chinese geological entity spatial relationship corpus and proposes a joint extraction model (GeoESRJE) for geological entity spatial relations. This method regards the two subtasks of entity recognition and relationship extraction as a sequence labeling task, realizes mutual promotion between subtasks, reduces error propagation and information loss, and improves the ability to extract geological entity spatial relations. The spatial relations in the corpus are divided into three categories (topological relations, absolute direction relations, and relative direction relations) and 22 subcategories. We used the test data set to validate the proposed model, and compared with mainstream models, GeoESRJE showed improvements across various evaluation metrics, confirming the effectiveness of the proposed method. Based on the test results, we constructed a knowledge graph of spatial relationships extracted from different themes. By providing a detailed description of these spatial relationships, we laid an important foundation for the mining and analysis of geological knowledge.