OmEGa(<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" altimg="si1.svg" display="inline" id="d1e2076"><mml:mi>Ω</mml:mi></mml:math>): Ontology-based information extraction framework for constructing task-centric knowledge graph from manufacturing documents with large language model
Midan Shim, Hyojun Choi, Heeyeon Koo, Kaehyun Um, Kyong-Ho Lee, Sanghyun Lee
Abstract
Manufacturing industry relies heavily on technical documents that encapsulate specialized knowledge essential for optimizing production and maintenance processes. However, extracting meaningful insights from these documents is challenging due to their complex structure, domain-specific terminology, and multimodal content, which includes text, images, and tables. Furthermore, there is a contextual gap between the generic training data of pre-trained language models (PLMs) and the specialized knowledge required for manufacturing documents. To address these issues, a Task-Centric Ontology (TCO) is designed to describe fundamental manufacturing tasks, and develop OmEGa, an Ontology-based Information Extraction Framework for Task-Centric Knowledge Graphs . OmEGa leverages large language models (LLMs) to perform instance recognition and relation classification on multimodal documents. By utilizing spatial embedding and modality linking, OmEGa addresses structural challenges, while TCO-driven reasoning mitigates contextual challenges. Experimental results demonstrate the effectiveness of OmEGa, achieving strong performance on both proprietary and open-source datasets. Additionally, a Knowledge Graph Question Answering (KGQA) system built on the extracted task-centric knowledge shows promise in enhancing communication among domain experts in the manufacturing sector .