A Causal Relation Extraction Model for Power Fault Texts Integrating Causal Prompt Vectors

Aiming at the problems that causal relations in power fault texts have hidden expressions and complex structures, and existing models struggle to extract implicit causal relations, this paper proposes a causal relation extraction model named CPBERT (Causal Prompting-enhanced BERT), which combines a causal prompting mechanism and a gated fusion module. Firstly, a causal terminology table suitable for power operation and maintenance contexts is constructed, covering both explicit and implicit causal connectives, to generate prompt vectors that mark potential causal positions. Secondly, the prompt information and BERT-encoded context vectors are dynamically fused through a gated unit, enabling collaborative modeling of local causal semantics and global context vectors. Finally, the fused representation is input into the Softmax layer to complete causal relation extraction. Experiments based on fault logs from a provincial power system in China show that CP-BERT outperforms CNN, BiLSTM, and BERT in terms of Precision, Recall, and F1-score, with more significant improvements in the three metrics for implicit causal relation extraction. This verifies its effectiveness in causal relation extraction from power fault texts and provides a new technical pathway for causal knowledge mining in the power industry.