A Heuristic Grafting Strategy for Manufacturing Knowledge Graph Extending and Completion Based on Nature Language Processing: KnowTree

Abstract

Applied to search, question answering, and semantic web of close-or-open domain, knowledge graph (KG) is known for its incompleteness subject to the rapid knowledge growing pace. Inspired by the agricultural grafting technology to fruit variety, this paper proposes a heuristic knowledge grafting strategy (HGS) for manufacturing knowledge graph (MKG) named KnowTree extending and completion with natural language processing (NLP) mining engineering cases document. Based on similarity analysis, firstly the grafting related definitions and mechanisms (completeness, relatedness, connectivity and reutilization) are defined. Then, focused on the four mechanisms, HGS takes a pair same engineering documents as input. KnowWords is built as a collection of KnowScion, and each scion is mined from engineering documents based on the SAO structure network, whose importance is evaluated by SAORank counting the in-out degree of centrality. On another hand, the KnowRoot system is designed based on the extended P & S ontology model to characterize the structure of abstract document into four sub-space of knowledge: know-what (problem), know-why (context), know-how (solution) and know-with (result), where a pre-trained language representation model K-BERT is used to classify the KnowScion candidates into the designed KnowRoot system with a fine-tuning classification task. In the knowledge grafting process, the connection unit is constructed based on the extracted domain knowledge triples of the K-BERT model, where the head element of a triple is from the KnowScion candidate set KnowWords satisfying the threshold value, the tail element is from the domain MKG to be extended, and a connection factor is used to evaluate the relationship of union combination. To the goal of knowledge reuse, the path based reasoning rules are designed for KnowTree reutilization. Finally, take the latest engineering case abstract (ECA) in whitegoods domain as resources, a case study is conducted to validate the proposed HGS strategy.

Keywords

• Knowledge graph extending and completion
• heuristic grafting strategy (HGS)
• NLP