Nowadays, description-based approaches are widely employed for the link prediction task on knowledge graphs, largely due to the integration of pre-trained language models (PLMs). These approaches enable the effective exploration and utilization of the rich textual data present in knowledge graphs. However, despite their strengths, models that rely on textual feature descriptions still face certain limitations compared to structure-based methods, particularly in terms of providing additional contextual information for entities. Structure-based methods, by representing entities and their relationships in embedding spaces, deliver stronger performance by allowing entities to carry contextual information through their relative positions to neighboring entities. In this paper, we propose the Struct-KGS2S model for link prediction on knowledge graphs, which integrates structural information from the graph with a seq2seq PLM-based model. To assess our model, we conducted experiments on three widely-used link prediction datasets: FB15k-237, WN18RR, and Wikidata5M, evaluating performance using the MRR and Hit@K metrics. Our experimental results demonstrate that the proposed method not only surpasses existing PLM-based models on these metrics but also achieves competitive results with structure-based models, setting new state-of-the-art performance on several benchmark datasets.
