Electrical Equipment Knowledge Graph Embedding Using Language Model with Self-learned Prompts

Abstract

In recent years, knowledge graphs have had a significant impact across diverse domains. Notably, the power knowledge graph has garnered considerable attention as high-performance database. However, its untapped reasoning capabilities offer an enticing avenue for exploration. One of the main reasons is the sparsity of power grid datasets, especially electrical equipment knowledge graph. Because of the scarcity of high-risk records, there exists a large number of long tail entities and long tail relations. To address this challenge, we introduce a novel text-based model called GELMSP (Graph Embedding using Language Model with Self-learned Prompts). We employ a bi-encoder structure along with a contrastive learning strategy to expedite the training process. Additionally, our approach incorporates a self-learned prompt mechanism that generates prompts for specific situations without the need for any additional information, known as self-learning. This harnesses the power of pre-trained language models to comprehend the semantic nuances within the entities and relationships of the knowledge graph. Adopting this innovative method enables our model to effectively handle sparse datasets, leading to a comprehensive understanding of interconnectedness within the knowledge graph. Additionally, we demonstrate the efficacy of our model through extensive experiments and comparisons against baseline methods, reaffirming its potential in advancing the state-of-the-art in electrical equipment defect diagnosis.