Enriched entity representation of knowledge graph for text generation

Abstract Text generation is a key tool in natural language applications. Generating texts which could express rich ideas through several sentences needs structured representation of their content. Many works utilize graph-based methods for graph-to-text generation, like knowledge-graph-to-text generation. However, generating from knowledge graph still faces problems, such as repetitions and the entity information not fully utilized generated text. In this paper, we focus on develop multi-level fusion (MEFR) model to address above aiming generate high-quality text graph. Our introduces mechanism, capable aggregating node representations word level phrase obtain Then, Graph Transformer adopted encode outputs contextualized representations. Besides, vanilla beam search-based comparison mechanism during decoding procedure, further considers similarity reduce repetitive Experimental results show that proposed MEFR effectively improve performance, outperform other baselines AGENDA WebNLG datasets. The also demonstrate importance explore contained