Abstract
Geoscience data associated with mineral resource surveys have become essential digital assets for governments and mining companies. The rapid increase in the volume of geoscience data makes it challenging to acquire knowledge quickly. In this study, we proposed and built a workflow that employs knowledge graph techniques, deep learning, question templates, and matching algorithms to provide a lightweight question-answering service for field-based geologists involved in mineral resource surveys. Initially, we utilized deep-learning-based geological entities and their semantic relation recognition, along with relational data mapping, to construct the mineral resource survey knowledge graph based on the ontology model. We then employed question template matching, a geological entity recognition model, and a sentence transformer to determine the optimal question template and generate a query statement for knowledge acquisition from a knowledge graph based on the Cypher language. Subsequently, we utilized a subgraph and a short abstract to express the results. The comparison with large language models and retrieval-augmented generation indicates that our solution is suitable for field-based mineral source surveys in a poor network environment with low-performance devices, data privacy concerns, and narrowly focused topics. The results also suggest that further studies on geoscience pre-trained models, an informative library of question templates, and multimodal knowledge graphs are necessary to improve the performance of the knowledge graph-driven question-answering system.