Despite significant popularity, Large-language Models (LLMs) require explicit, contextual facts to support domain-specific knowledge-intensive tasks in the design process. The applications built using LLMs should hence adopt Retrieval-Augmented Generation (RAG) to better suit the design process. In this article, we present a data-driven method to identify explicit facts from patent documents that provide standard descriptions of over 8 million artefacts. In our method, we train roBERTa Transformer-based sequence classification models using our dataset of 44,227 sentences and facts. Upon classifying tokens in a sentence as entities or relationships, our method uses another classifier to identify specific relationship tokens for a given pair of entities so that explicit facts of the form head entity :: relationship :: tail entity are identified. In the benchmark approaches for constructing facts, we use linear classifiers and Graph Neural Networks (GNNs) both incorporating BERT Transformer-based token embeddings to predict associations among the entities and relationships. We apply our method to 4,870 fan system related patents and populate a knowledge base of around 3 million facts. Upon retrieving the facts representing generalisable domain knowledge and the knowledge of specific subsystems and issues, we demonstrate how these facts contextualise LLMs for generating text that is more relevant to the design process.

翻译：尽管大语言模型（LLMs）广受欢迎，但在设计过程中，它们仍需显式、上下文相关的事实来支持领域特定的知识密集型任务。基于LLMs构建的应用应采用检索增强生成（RAG）以更好地适配设计流程。本文提出一种数据驱动方法，从涵盖800余万件制品标准化描述的专利文档中识别显式事实。在该方法中，我们使用包含44,227条句子与事实的数据集训练基于roBERTa Transformer的序列分类模型。通过将句子中的词元分类为实体或关系，我们的方法利用另一个分类器针对给定实体对识别特定关系词元，从而提取形如"头实体::关系::尾实体"的显式事实。在构建事实的基准方法中，我们使用线性分类器与图神经网络（GNNs），两者均结合基于BERT Transformer的词元嵌入来预测实体与关系间的关联。我们将该方法应用于4,870项风机系统相关专利，构建了包含约300万条事实的知识库。通过检索代表可泛化领域知识的事实以及特定子系统与问题的知识，我们展示了这些事实如何为LLMs提供上下文，从而生成与设计流程更相关的文本。