A Domain Knowledge–Enhanced Theme Expansion Method for Earth Surface System Science Dataset Recommendation

ABSTRACT With the advancement of Earth observation technologies, open scientific datasets for the Earth surface system have experienced exponential growth. However, traditional retrieval methods face the challenge of low recall due to the significant vocabulary mismatch between users' natural language queries and dataset metadata. Recently, generative query expansion methods based on large language models (LLMs) have demonstrated promising application prospects. Nevertheless, in the highly specialized geoscience domain, the lack of knowledge constraints makes LLMs highly susceptible to generating “hallucinations” and introducing semantic noise, thereby degrading retrieval quality. To address these issues, this paper proposes a domain knowledge‐enhanced theme expansion method for Earth surface system science dataset recommendation. First, a hybrid construction strategy is employed to establish a domain knowledge base featuring a three‐tier “domain‐theme‐instance” structure, providing a controlled semantic space for the LLM's query expansion. Second, a controlled query expansion mechanism is proposed, transforming the LLM from a “free text generator” into a “constrained semantic selector.” When processing a user query, the LLM navigates research intents within the knowledge base and selects instance‐level thematic keywords that genuinely exist in the metadata for expansion, thereby effectively suppressing semantic noise. Experimental results on a self‐constructed geoscience evaluation dataset (GeoFinder) and an open‐source machine learning benchmark dataset (DataFinder) demonstrate that the proposed method achieves significant improvements in both recall and ranking quality (NDCG, MRR) under sparse and dense retrieval architectures, exhibiting excellent cross‐domain generalization capability and robustness.