Semantic topic models for source code analysis

Abstract

Topic modeling techniques have been recently applied to analyze and model source code. Such techniques exploit the textual content of source code to provide automated support for several basic software engineering activities. Despite these advances, applications of topic modeling in software engineering are frequently suboptimal. This can be attributed to the fact that current state-of-the-art topic modeling techniques tend to be data intensive. However, the textual content of source code, embedded in its identifiers, comments, and string literals, tends to be sparse in nature. This prevents classical topic modeling techniques, typically used to model natural language texts, to generate proper models when applied to source code. Furthermore, the operational complexity and multi-parameter calibration often associated with conventional topic modeling techniques raise important concerns about their feasibility as data analysis models in software engineering. Motivated by these observations, in this paper we propose a novel approach for topic modeling designed for source code. The proposed approach exploits the basic assumptions of the cluster hypothesis and information theory to discover semantically coherent topics in software systems. Ten software systems from different application domains are used to empirically calibrate and configure the proposed approach. The usefulness of generated topics is empirically validated using human judgment. Furthermore, a case study that demonstrates thet operation of the proposed approach in analyzing code evolution is reported. The results show that our approach produces stable, more interpretable, and more expressive topics than classical topic modeling techniques without the necessity for extensive parameter calibration.