Domain knowledge-guided geological named entities recognition of rock minerals based on prompt engineering with error feedback mechanism

Abstract

Geological reports contain a wealth of information about geological entities, such as rocks and minerals, which are of great significance for resource exploration, environmental assessment, 3D geological modeling, and intelligent prospecting. However, existing methods and models (e.g., deep learning-based approaches) for geological named entity recognition (GNER) heavily rely on large manually annotated corpora. This process is time-consuming and labor-intensive, and it faces limitations when dealing with complex entities in geological reports, such as long or nested entities. To address this issue, this paper proposes a rock and mineral NER method based on prompt engineering and domain knowledge guidance. First, preliminary entity recognition is conducted through labeling rather than extraction, mitigating the problem of repetitive recognition of nested entities. Second, we summarize and categorize the types of errors made by large language models(LLMs), incorporating geological knowledge guidance for secondary recognition to reduce common mistakes. Finally, secondary category validation is used to alleviate the “hallucination” problem, where LLMs mistakenly identify non-entities as entities. This method requires only two examples as training samples to guide the model, significantly reducing the workload of corpus annotation. Experiments were conducted on multiple LLMs (e.g., GPT-4o-0513, GPT-4o-0806, and Claude 3.5 sonnet). The results show that on our self-constructed dataset, compared to direct entity extraction, the accuracy of rock and mineral recognition is improved by approximately 17 % and 11 %, respectively, validating the effectiveness of combining domain knowledge with LLMs for GNER.