Abstract: Aiming at the problems that the  knowledge graph completion method based on  Transformer as the backbone network, included parameter redundancy in feed-forward networks, difficulties in identifying tail entities under commonsense scenarios, and embedding biases in contrastive learning, we proposed an adapter-enhanced knowledge graph completion algorithm that 
integrated large language models with multi-positive sample contrastive learning. The algorithm reduced redundant features by introducing multi-head adapters in feed-forward network, and utilized  large language models to enhance commonsense reasoning ability. At the same time, it corrected embedding biases through  multi-positive sample contrastive learning. Experimental results 
show that, compared to the current state-of-the-art models, the algorithm improves MRR by 5.4% and 9.2% on WN18RR and FB15k-237 datasets, respectively, and by  3.6% and 6.7% in transductive and inductive settings  on the more complex Wikidata5M dataset,  respectively, and demonstrates superior generalization ability under low-resource and complex scenarios.

Key words: knowledge graph completion, knowledge graph, large language model, contrastive learning, adapter learning