CEDG-GeoQA: Knowledge base question answering for the geoscience domain via Chinese entity description graph

• Published in: Earth science informatics Vol. 17; no. 3; pp. 2609 - 2621
• Main Authors: Wei, Lai, Lu, Qinghua, Duan, Yilin, Yao, Hong, Kang, Xiaojun
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.06.2024; Springer Nature B.V
• Subjects: Earth and Environmental Science; Earth science; Earth science research; Earth Sciences; Earth System Sciences; Geography; Graph theory; Information Systems Applications (incl.Internet); Knowledge bases (artificial intelligence); Natural language; Ontology; Queries; Questions; Scientific research; Search engines; Simulation and Modeling; Space Exploration and Astronautics; Space Sciences (including Extraterrestrial Physics; Semantic parsing; Geoscience domain; Geoscience question answering; Knowledge graph
• ISSN: 1865-0473; 1865-0481
• DOI: 10.1007/s12145-024-01304-8

Acquiring geoscience knowledge is crucial for advancing earth science research. Currently, geoscience knowledge can be obtained through search engines or specialized databases. However, the quality of search engine results varies, and geoscience databases do not support natural language queries. To address these challenges, Geoscience Question Answering (GeoQA) systems have been developed to provide answers to natural language queries. Much of the existing research in geoscience QA primarily focuses on geography, with other domains remaining relatively unexplored. To bridge this gap, our study introduces a Chinese geoscience QA dataset that covers a wide range of topics, including geography, climate, and culture. Additionally, we propose the CEDG-GeoQA framework for Chinese geoscience QA. The model begins by utilizing syntactic parsing to convert unstructured queries into an entity description graph (EDG). Subsequently, it aligns the EDG with a comprehensive geoscience knowledge base, extracting a subgraph centered around the subject entity. This subgraph is used to assess candidate answers and determine the most likely response. Our comprehensive experiments, conducted using a Chinese geo-knowledge base, demonstrate the superior performance of our method, achieving a 5% improvement in the F1 measure compared to existing baselines, including WDAqua, gAnswer, and NSQA.