Physical model building is essential for realizing digital twins in the manufacturing industry and requires much toil. We aim to develop automated physical model builder (AutoPMoB) that can automatically build physical models from literature databases. AutoPMoB requires several fundamental technologies, and domain-specific datasets play a vital role in developing such technologies. Although datasets related to variables have been created, there has been no dataset in the chemical engineering domain. To create such a dataset, in this study, we developed an algorithm for extracting variable symbols from documents and a variable annotation tool, VARAT, based on the algorithm. We used the tool and created a dataset containing about 1,733 variable symbols from 45 papers on physical models of five manufacturing processes. VARAT enables us to quickly and accurately extract the variable symbols from documents and reduces the time for annotation per paper to less than half, which streamlines the annotation process.

Mathematical formulas are essential tools for conveying mathematical concepts. Definitions of symbols in mathematical formulas often vary among different documents; thus, knowing the definitions is fundamental to grasping the semantics of the formulas. This research targets how to extract definitions of symbols representing variables from documents on chemical processes. We defined three features focusing on the unique usage of variable symbols and definitions in these documents and proposed a new variable definition extraction method. We compared the performance of the proposed method with that of a representative conventional method using 45 papers on five chemical processes. The proposed method achieved higher accuracy than the conventional one for four processes. We also demonstrated that our newly defined features contributed to the performance improvement and that the proposed method can achieve high accuracy with a small number of training datasets.