Recently, pre-trained language models (PLMs) have greatly advanced the Knowledge Graph-to-Text (KG-to-Text) tasks. However, the structure information of knowledge graph is still underexplored, and certain entities in the knowledge graph are often omitted in generated text. Even the powerful capabilities of LLMs are not well utilized. To address these issues, this paper proposes an integrated three-stage framework (ITS) for enhanced KG-to-Text generation, which leverages improved graph encoder, Mention-Flag guided decoder, and LLM-Powered post-processor. In the knowledge graph encoding stage, this paper proposes a Bidirectional View Graph Encoder (BVGE) based on GNN, which employs a bidirectional view encoding method specifically designed for encoding complex knowledge graphs. We then combine the BVGEs structural encoding with the semantic encoding of pre-trained models via a fusion module as the knowledge graph representation. In the decoding stage, we design a Mention-Flag guided decoder to avoid missing the entities in KGs. In the post-processing stage, we used Llama3 to correct and refine the generated text. Results on multiple benchmark datasets show that our framework significantly outperforms state-of-the-art methods.
Liu et al. (2026) studied this question.