GeoSMIE: An event extraction framework for Document-Level spatial morphological information extraction

Spatial morphological information (SMI) in geological texts provides critical insights into the formation, localization, and distribution of geological bodies. However, SMI is often scattered across multiple sentences or coexists in complex forms within the same document, making it challenging to extract using traditional methods. In this paper, we address this gap by formalizing SMI extraction as an event extraction task and proposed a novel Geological body SMI Extraction model, GeoSMIE. Our approach is innovative in two key ways: first, we implement a no-trigger-word annotation strategy to capture both descriptive and digital SMI, ensuring that SMI without explicit morphological triggers is not missed. Second, we design dual graph neural networks (GNNs) to handle long-distance dependencies and complex interactions between scattered arguments across sentences. To validate its effectiveness, we compared GeoSMIE to state-of-the-art models on the constructed dataset. For SMI extraction, GeoSMIE outperformed the optimal baseline by 0.4%, 2.2%, and 1.5% for accuracy, recall, and MicroF1 score, respectively. This work provides an innovative idea for extracting complex spatial information from geoscience texts.