ProcessBERT: Towards Equivalence Judgment of Variable Definitions among Multiple Engineering Documents

Abstract

Physical models play an important role in the process industry. However, conventional physical model building requires a survey on a huge amount of literature and trial-and-error to improve the model performance. We aim to develop an automated physical model builder (AutoPMoB), which automatically collects documents about a target process from literature databases, extracts necessary information from them, and builds a desired physical model by reorganizing the information. In this study, we proposed a method of judging equivalence of variable definitions, which is one of the fundamental technologies to realize AutoPMoB. We built a large-scale corpus specialized in chemical engineering and developed ProcessBERT, which is a domain-specific language model pre-trained on our corpus. We created datasets from papers related to chemical processes and evaluated the performance of ProcessBERT in the equivalence judgment task. We found that ProcessBERT outperformed the other language models in the similarity-based method.

Tasks

Reproductions