Understanding ore-forming conditions using machine reading of text

Abstract

• A methodology is proposed for constructing a geological knowledge graph (KG) from mineral exploration reports. • KG is enriched using an external stratigraphic database to incorporate important relations including hierarchical relations and name changes among extracted stratigraphic units. • Sequence labelling is used for geological named entry recognition, while word semantics, distributed representations and clustering are used to extract relation types. • Case studies on mineral deposits are used to demonstrate the feasibility of the proposed methodology. Mineral exploration reports assist present day exploration by providing valuable observations about the geological environments in which mineral deposits form. Querying and aggregating historical data can assist in reducing future exploration risks and costs. However, since the reports are written in unstructured text, it is a challenging task to derive meaningful geological information without manually reading through a large collection of reports, which is a formidable task for geologists. In this study, geological information relevant to mineralisation and ore-forming conditions is automatically extracted from such under-utilised exploration reports. This is achieved by constructing knowledge graphs that describe geological entities and their relations as they appear in exploration reports. Natural language processing and deep learning methods are used to automatically extract and label geological terms with the correct entity types and establish the relationships between these entities. In this research, six dominant entity types are considered, namely, geographical location, geological timescale, stratigraphic unit, rock type, ore and deposit type, and contained minerals. Two knowledge graphs are constructed for two high-quality mineral exploration reports, one for iron ore and the other for gold deposit, to illustrate the effectiveness of our methodology. The knowledge graphs are then assessed by determining whether the contents of the source reports were depicted accurately, specifically the labelled geological terms in the nodes and their associations in node connections. The results show that the structured information stored in the knowledge graphs faithfully represent the contents of the source reports, matching well with the domain knowledge. The proposed methods are capable of rapidly and robustly transforming text data into a structured form, the untapped area towards geological knowledge mining.