Variable Extraction from Scientific Papers for Automatic Physical Model Building

Abstract

Physical models are used to design chemical plants, to improve an operation condition, and so on. The information about existing physical models can be obtained from papers, and a more accurate model can be built by improving or combining past models. Although the past models are useful for building a new model, it takes too much effort and time to find out useful models accurately from papers. The goal of this research is to develop the automatic physical model building system, which proposes the best-desired model. The information about the variables in the model is essential for model building. For example, when we want to build the model which includes variables a, the past model, which does not include the variables a, would not be useful. Most past researches aim to extract the meaning of the equation and the equation itself; however, we cannot extract all of the variables and the meaning of variables. Also, most of the past variable extraction methods are based on supervised learning and require labeled data. Since the physical models are proposed in lots of areas, we cannot create labeled data throughout those areas. In this research, we proposed a method to extract the definition of paper variables based on the characters used as the variables, subscripts, and superscripts.