Discovering Important Source Code Terms

Terms in source code have become extremely important in Software Engineering research. These ``important'' terms are typically used as input to research tools. Therefore, the quality of the output of these tools will depend on the quality of the term extraction technique. Currently, there is no definitive best technique for predicting the importance of terms during program comprehension. In my work, I perform a literature review of several techniques. I then propose a unified importance prediction model based on a machine learning algorithm. I evaluate my model in a field study involving professional programmers, as well as a standard 10-fold synthetic study. I found that my model predicts the top quartile of most-important source code terms with approximately 50\% precision and recall, outperforming tf/idf and other popular techniques. Furthermore, I found that, during actual program comprehension tasks, the predictions from my model help programmers equivalent to a real set of most-important terms.