MGSA: Multi-Granularity Graph Structure Attention for Knowledge Graph-to-Text Generation

For the knowledge graph to text(KG-to-text) generation task, recent works have attempted to incorporate graph structure information into pre-trained language models(PLMs) to capture the structure information of knowledge graphs. However, these improvements only capture single-granularity structure information, either between entities in the original graph or between words within or across entities. Considering only entity-level structures neglects the semantic information between words, while focusing solely on word-level structures ignores the relationships between complete entities. Therefore, this paper proposes the Multi-Granularity Graph Structure Attention (MGSA), which integrates both granularities of information. The encoder of the model architecture features an entity-level structure encoding module, a word-level structure encoding module, and an aggregation module that synthesizes information from both structure. This allows the model to more comprehensively understand the information contained in the original knowledge graph, thereby improving the quality of the generated text. We evaluated the MGSA model on two popular KG-to-text generation benchmark datasets and achieved superior performance compared to models based on single-granularity structures.