A two-stage knowledge graph completion based on LLMs’ data augmentation and atrous spatial pyramid pooling

With the development of information technology, a large amount of unstructured and fragmented data is generated. Knowledge graphs can effectively integrate these fragmented data. Due to the difficulty of domain knowledge mining, knowledge graphs have problems of data sparseness and data missing. In addition, standard convolutional neural networks have limited capability in capturing feature interactions. To address data sparsity and the limitations of standard convolutional models, we propose DA-ARKGC, a two-stage knowledge graph completion model using wheat as a case study. In the first stage, to address the data sparsity problem, the rule mining data augmentation module (DA) based on large language models expands the wheat knowledge graph. In the second stage, the knowledge completion module (ARKGC) of the atrous spatial pyramid pooling with residual is introduced to achieve knowledge completion. The DA-ARKGC model was verified on the constructed wheat knowledge graph (Wheat_KG). Compared with ConvE, its MRR, Hits@1, Hits@3 and Hits@10 increased by 10% and 10.2%, 10.1% and 9.3%, respectively. In order to verify the effectiveness and generalization of the ARKGC module, experiments were conducted on the open-source datasets WN18 and FB15k. The results demonstrated that the model achieved optimal or sub-optimal performance compared with other baseline models.