Abstract
Natural language artefact descriptions are the primary carriers of engineering design knowledge, acting as outcomes as well as drivers of the design process. Engineering design literature aims to support knowledge retrieval, representation, and reuse to facilitate various design tasks. Traditional perspectives of design research offer ontological schemes for capturing certain facets of design knowledge, while often under-representing the explicit design knowledge that is communicated in artefact descriptions. Contemporary design research tends to leverage Large-Language Models (LLMs) that require contextualising with explicit design knowledge to produce technically appropriate responses that are useful in a design environment.
In my thesis, I address these issues by explicating and examining design knowledge to understand its basis. To explicate design knowledge, we develop an innovative approach for relation extraction from patented artefact descriptions, by creating a dataset of 375,084 examples from 44,227 sentences in 4,205 patents. We train various transformers for relation identification and elicitation, achieving up to 99.7% accuracy.
To examine design knowledge, we organise over 24 million facts identified from 33,881 patented artefact descriptions into artefact knowledge graphs. Zipf distribution analyses reveal that over 50% of entities and relationships could be generalized to 64 and 24 syntaxes that represent the linguistic basis of design knowledge. The relationships largely represent attributes, structure, purpose, hierarchy, exemplification, and behaviour. Motif analyses on the graph structures reveal the 8 most dominant subgraph patterns representing the structural basis of design knowledge, which could be simplified into sequence, aggregation, and hierarchy.
Based on these results, I propose a few regulatory precepts for concretizing entities and relationships and explicating hierarchical structures, potentially aiding construction and modularisation of design knowledge. Overall, the research outcomes offer significant contributions, including a relation extraction method that is useful for Retrieval-Augmented Generation (RAG) with LLMs and insights into the influence of structural properties on knowledge graphs. These findings can enhance knowledge retrieval, representation, and reuse in design research, education, and practice, and aid knowledge management systems.