209-8 Ontology-Guided Construction of Geoscience Knowledge Graphs Using Large Language Models

Session: Deep-Time Earth and the AI Revolution

Presenting Author:

Grant Boquet

Authors:

Yu, Ting¹, Feng, Zhiqiang², Jiang, Ting³, Wu, Wenxun⁴, MOHAMED, Jaward Bah⁵, Chen, Hongyang⁶, Ye, Jieping⁷, Boquet, Grant Michael⁸, Hu, Huiting⁹, Li, Yilin¹⁰, Tang, Xianming¹¹

(1) Zhejiang Lab, Hangzhou City, Zhejiang Province, China, (2) Zhejiang Center for Deep-time Digital Earth, Hangzhou City, Zhejiang Province, China; Northeast Petroleum University, Daqing, Heilongjiang Province, China, (3) Zhejiang Lab, Hangzhou City, Zhejiang Province, China, (4) Zhejiang Lab, Hangzhou City, Zhejiang Province, China, (5) Zhejiang Lab, Hangzhou City, Zhejiang Province, China, (6) Zhejiang Lab, Hangzhou City, Zhejiang Province, China, (7) Zhejiang Lab, Hangzhou, Zhejiang Province, China, (8) GeoGPT, Zhejiang Lab, , (9) Northeast Petroleum University, Daqing, Heilongjiang Province, China, (10) Northeast Petroleum University, Daqing, Heilongjiang Province, China, (11) Petroleum Exploration and Production Research Institute, SINOPEC, Beijing, China,

Abstract:

Knowledge graphs serve as foundational semantic infrastructures for integrating heterogeneous geoscience data—such as geological, mineral, and environmental information—and are essential for knowledge-driven intelligent analysis and decision support. However, traditional knowledge graph construction faces challenges in extraction automation, semantic consistency, and ontology alignment. To address these issues, this paper proposes a method for constructing geoscience knowledge graphs by combining large language models (LLMs) with ontological constraints, enabling high-precision, structurally sound, and semantically interpretable automated knowledge extraction.

The method employs predefined geoscience ontologies as structural and semantic frameworks, embedding their classes, attributes, relational hierarchies, and logical rules into LLM prompt designs to guide knowledge extraction. Structured prompts enable LLMs to identify ontology-compliant entities and relationships from unstructured texts—including research literature, technical reports, and field records—effectively reducing hallucinations and improving semantic accuracy. An iterative verification mechanism compares initial extractions with the ontology via logical reasoning to detect and correct invalid triples. Few-shot learning and chain-of-thought prompting further enhance model performance in low-resource settings, while rule-based validation ensures consistency and reduces noise.

Evaluations on real-world geoscience text datasets show that, compared to non-ontology-guided baselines, the proposed approach improves relationship extraction accuracy by 22.6% and achieves an F1 score of 84.3%. The resulting knowledge graph demonstrates superior interpretability and ontological coherence. This study presents semantic guidance and generative intelligence paradigm for high-quality domain-specific knowledge graph construction, advancing geoscience knowledge management toward automation, standardization, and intelligence.

Geological Society of America Abstracts with Program. Vol. 57, No. 6, 2025

doi: 10.1130/abs/2025AM-11110

© Copyright 2025 The Geological Society of America (GSA), all rights reserved.