TLF ‐Net‐Edge: A Lightweight Text‐Enhanced Local Structure‐Aware Network for English Knowledge Graph Embedding on Edge Intelligence Networks

ABSTRACT

The distributed deployment characteristics and low resource constraints of edge intelligent networks make English Knowledge Graph Embedding (KGE) face core problems such as high inference latency, semantic accuracy loss, and poor resource adaptability. Traditional text‐structure fused KGE models are difficult to deploy directly due to their complex structures and redundant parameters. To address this issue, a lightweight text‐enhanced local structure‐aware network, TLF‐Net‐Edge, is proposed to balance the efficiency and performance of English KG representation learning in edge scenarios. This model innovatively integrates the text‐local structure fused KGE with the edge‐cloud collaborative architecture for the first time: it designs a lightweight semantic disambiguation module adapted to the English context to simplify the model while retaining key linguistic features such as word roots and parts of speech; constructs a semantically guided local structure subgraph to reduce the computational overhead at the edge; and proposes an edge‐oriented dynamic balance strategy to achieve dynamic matching among performance, latency, and resource consumption through resource‐aware inference mode switching and adaptive feature fusion. Experimental results on the FB15K‐237 and DB15K English KG datasets show that TLF‐Net‐Edge significantly outperforms baseline models such as TransE, ConvE, and KG‐BERT in core metrics including Mean Reciprocal Rank (MRR) and Hits@N. Benefiting from the edge‐cloud collaborative design, the model reduces inference latency by 89.5% and memory occupancy by 91.2% compared with TGforme model. Our model parameter size is compressed to less than 5 million, and the inference latency is reduced by 78.2% compared with the original TLF‐Net, as demonstrated in the ablation experiments. This study provides key technical support for the application of English KGs in scenarios such as edge intelligent question answering and real‐time academic data analysis, and expands the application boundary of KGE models in edge intelligent networks.