Fact-condition statements and super relation extraction for geothermic knowledge graphs construction

Researchers utilize information from the geoscience literature to deduce the regional or global geological evolution. Traditionally this process has relied on the labor of researchers. As the number of papers continues to increase, acquiring domain-specific knowledge becomes a heavy burden. Knowledge Graph (KG) is proposed as a new knowledge representation technology to change this situation. However, the super relation is not considered in the previous KG, which bridges the geological phenomenon (fact) and its precondition (condition). For instance, in the statement (“the late Archean was a crucial transition period in the history of global geodynamics”), the condition statement (“crucial transition for global geodynamics”) works as the complementary fact statement (“the late Archean was a crucial transition period”), which defines the scale of crucial transition accurately in the late Archean. In this study, fact-condition statement extraction is introduced to construct a geological knowledge graph. A rule-based multi-input multi-output model (R-MIMO) is proposed for information extraction. In the R-MIMO, fact-condition statements and their super relation are considered and extracted for the first time. To verify its performances, a GeothCF dataset with 1455 fact tuples and 789 condition tuples is constructed. In experiments, the R-MIMO model achieves the best performance by using BERT as encoder and LSTM-d as decoder, achieving F1 80.24% in tuple extraction and F1 70.03% in tag prediction task. Furthermore, the geothermic KG with super relation is automatically constructed for the first time by trained R-MIMO, which can provide structured data for further geothermic research.