Learning I/O Variables from Scientific Software's User Manuals

Scientific software often involves many input and output variables. Identifying these variables is important for such software engineering tasks as metamorphic testing. To reduce the manual work, we report in this paper our investigation of machine learning algorithms in classifying variables from software's user manuals. We identify thirteen natural-language features, and use them to develop a multi-layer solution where the first layer distinguishes variables from non-variables and the second layer classifies the variables into input and output types. Our experimental results on three scientific software systems show that random forest and feedforward neural network can be used to best implement the first layer and second layer respectively.