Engineering Knowledge Graph Construction from Technical Documents via OpenIE and T5-driven Hybrid Extraction

Abstract

The automatic extraction of structured engineering knowledge from unstructured technical papers is the second requirement for computer-aided engineering intelligence. In order to create a general-purpose engineering knowledge graph from a variety of text resources, this study presents a novel combination of Open Information Extraction (OpenIE) and Transformer-based semantic modeling (T5). First of all, the domain jargon and intricate technical syntax can be handled by carefully prepared preprocessing and token normalization. A T5 semantic encoder uses deep contextual representations to further disambiguate context and increase precision once OpenIE modules have extracted candidate entity-relation triples. The suggested approach outperforms the best-performing baselines, as demonstrated by experimental evaluation on a hierarchically annotated engineering corpus; it has a macro F1 score of 0.82 and a micro F1 score of 0.85, improving by 7.3% and 6.8%, respectively. Over 65% of the net performance boost is attributed to transformer-based semantic filtering, according to ablation studies. Coreference ambiguity and relation type overlap are the primary residual issues, according to error analysis. Over 90% of the original accuracy has been maintained, and several testings in various fields have confirmed that this approach can be applied in others. According to the aforementioned analysis, this approach offers great support for next technical document intelligence research and has good scalability and accuracy for the management of engineering information.