A General Software Readability Model

We present a generalizable formal model of software readability based on a human study of 5000 participants. Readability is fundamental to maintenance, but remains poorly understood. Previous models focused on symbol counts of small code snippets. By contrast, we approach code as read on screens by humans and propose to analyze visual, spatial and linguistic features, including structural patterns, sizes of code blocks, and verbal identifier content. We construct a readability metric based on these notions and show that it agrees with human judgments as well as they agree with each other and better than previous work. We identify universal features of readability and languageor experience-specific ones. Our metric also correlates with an external notion of defect density. We address multiple programming languages and different length samples, and evaluate using an order of magnitude more participants than previous work, all suggesting our model is more likely to generalize.