Fusing graph structural information with pre-trained generative model for knowledge graph-to-text generation

Knowledge graph-to-text generation (KG-to-Text) is a task that involves generating accurate textual descriptions based on a given knowledge graph. Previous efforts have often enhanced pre-trained generative models by incorporating additional auxiliary pre-training tasks to supplement the missing graph structural information. However, such tasks not only require substantial computational resources, but also yield limited improvements. To address this issue, we propose a novel approach that effectively combines the graph structural information from knowledge graphs with pre-trained generative models without altering their fundamental architecture. Our approach involves inputting the original knowledge graph data into a graph convolutional network to acquire graph feature representations enriched with node characteristics. Additionally, the linearized sequence of the knowledge graph is inputted into the pre-trained generative model to exploit its inherent semantic richness. After computing multi-head attention mechanisms, we fuse the acquired graph feature representations into the pre-trained generative model to supplement the missing graph structural information. Experimental results on the WebNLG and EventNarrative datasets show that our approach achieves improved performance while reducing computational overhead.