기계독해를 통한 설비 사양 문서 정형화

Abstract

Machine Reading Comprehension (MRC) is a task that entails extracting information by lever- aging a machine's language understanding capabilities. In the context of this study, our primary objective is to adeptly process extensive volumes of unstructured document data using MRC, thereby elevating the precision of decision-making processes. Our innovative approach intro- duces the Multi-Span MRC model, specifically tailored to tackle the complexities of Multi-Span extraction tasks that posed challenges for the conventional Single-Span MRC approach. The methodology at the core of our research places emphasis on the extraction of Part (P), Property (R), and Value (V) information from design specification documents, all within a single pas- sage. Through meticulous comparisons with the outcomes of Single Span Masked MRC and BiLSTM-CRF models, our research robustly establishes the superior performance of the Multi- Span MRC model. Consequently, this study not only advances our understanding of MRC tech- nology but also provides valuable insights into its advantageous application within the plant industry.