LLM-based design document understanding for semantic knowledge extraction and design verification

Abstract

Design documents contain essential design knowledge such as designers’ intent, decision-making criteria, and constraints, and are widely used to support accurate and consistent product development. Most design documents are extensive and composed of unstructured natural language-based text, which makes it difficult for humans to manually read and retrieve relevant information efficiently. To address this challenge, this study proposes an LLM-based design document understanding approach for structurally organizing implicit design information embedded in unstructured design documents. A sentence-level semantic classification framework is introduced to categorize design sentences according to quantitative and functional characteristics. In addition, contextual information is leveraged to identify the design object described by each sentence automatically. Design semantics classification and design object identification experiments were conducted by fine-tuning various Natural Language Processing (NLP) models and LLMs, achieving accuracy of 87.2% and 91.0%, respectively. Furthermore, a case study on automated design verification using real 3D CAD models demonstrates that the proposed approach can be directly applied to design automation tasks.