Efficient Incorporation of Knowledge Graph Information for Enhanced Graph-to-Text Generation

Previous research on Knowledge Graph-to-Text Generation (KG-to-Text) primarily introduced auxiliary pretraining tasks to enhance pre-trained generative models, aiming to address their limitations in handling graph structural information. However, this approach not only imposes substantial computational demands but also results in limited improvements. To address this issue, we propose an innovative method that effectively incorporates the structural information of knowledge graphs into pre-trained generative models without modifying their core architectures. Our approach involves inputting the original knowledge graph data into a graph convolutional network. Additionally, we feed linearized sequences derived from the knowledge graph into the pre-trained generative model to fully leverage its rich semantic information. By using a multi-head attention mechanism, we combine the obtained graph feature representations with the pre-trained generative model to address the model’s deficiencies in handling structural information. Experiment results on WebNLG and EventNarrative show that our approach not only reduces computational overhead but also achieves superior performance.