Enhancing RDF Verbalization with Descriptive and Relational Knowledge

RDF verbalization has received increasing interest, which aims to generate a natural language description of the knowledge base. Sequence-to-sequence models based on Transformer are able obtain strong performance equipped with pre-trained such as BART and T5. However, in spite general gain introduced by models, task is still limited small scale training dataset. To address problem, we propose two orthogonal strategies enhance representation learning triples. Concretely, types introduced, i.e., descriptive relational knowledge, respectively. The indicates semantic information self definition, learned from structural context. We further combine together learning. Experimental results WebNLG SemEval-2010 datasets show that can both model performance, their combination improvements most cases, providing new state-of-the-art results.