KGCDP-T: Interpreting knowledge graphs into text by content ordering and dynamic planning with three-level reconstruction

Abstract

Knowledge graph-to-text (KG-to-text) interpretation is employed to interpret given KG into semantically coherent and logically reasonable text to enhance the applicability of KG in more natural language generation (NLG) scenarios, such as search engines and text-dialog systems. Unfortunately, existing relevant research suffers from the semantic gap between structural knowledge graphs and unstructured text. To overcome this challenge, in this paper, we propose a novel pipeline-based Knowledge Graph interpreting model constructed by Content Ordering and Dynamic Planning with Three-Level Reconstruction (KGCDP-T). Specifically, the first “pipe” Content Ordering converts the given KG into several triple-groups and then into an ordered triple-sequence to plan the ordering by which the “content” in the given knowledge graph should be interpreted. Next, an entity-sequence is derived from the above ordered triple-sequence, where the second “pipe” Dynamic Planning captures the context shifting in the entity-sequence via Memory Network. In this way, the entity-sequence-level and memory-level contexts are learned and fused to generate the interpreted-text with more context-adaptive tokens. Moreover, the Three-Level Reconstruction mechanism is incorporated to capture the critical features transferred among the triple-groups, the ordered triple-sequence, the generated text-sequence and the given KG. The experimental results indicate that our proposed KGCDP-T can achieve the overall best KG-interpretation performance on content integrity (reflected by a 7.69% average improvement in Coverage), sentence fluency (reflected by 5.98% and 6.00% average improvements in BLEU and ROUGE-L, respectively) and logic coherency (reflected by 6.10% and 9.18% average improvements in METEOR and Chrf++, respectively) when compared with state-of-art KG-to-text models.