AddManBERT: A combinatorial triples extraction and classification task for establishing a knowledge graph to facilitate design for additive manufacturing

In recent years, triple extraction and classification have received attention in the context of Additive Manufacturing (AM). However, the lack of a formalized process to extract and classify triple from textual data poses challenges for the effective embedding learning techniques in utilizing AM's product innovation and manufacturing capabilities. Hence, the AM field's manual cognitive process hinders the broader adoption of Design for AM (DFAM) in manufacturing. Aiming to solve these challenging problems, this research proposes a Natural Language Processing (NLP) and Knowledge Graph (KG) methodology for triple extraction and classification from textual data to provide an embedding learning approach. Initially, multi-source textual data for triple extraction and classification is developed. Then, AM Bidirectional Encoder Representation from the Transformers (AddManBERT) is used for triple extraction and classification. The AddManBERT utilizes dependency parsing to determine the semantic relations between the entities for triple extraction and classification. Consequently, the AddManBERT transformed each extracted piece of knowledge from the textual data into a 768-dimensional vector structure by analyzing the projected probability of the output within the center word based on the token embedding surrounding the input. The triples extracted and classified are then saved in the Neo4j database and displayed as graph nodes. An experiment and an application case study verify the proposed method's efficacy. The experiment results indicate that the proposed method outperforms the traditional centralized approaches in responsiveness, classification accuracy, and prediction efficiency.