Hapax - Enriching Reverse Engineering with Semantic Clustering

Many reverse engineering approaches focus on structural information and ignore semantic information like the naming of identifiers or comments. But developers put their domain knowledge into exactly these parts of the source code. Without understanding the semantics of the code, one cannot tell its meaning. We use Latent Semantic Indexing, an information retrieval technique [3], to retrieve the semantic similarity between different entities (e.g. whole systems, classes and methods), and then we cluster these entities according to their similarity. We employ this technique to characterize entities by clustering their sub entities [2]. For example, we cluster classes to characterize the system. Furthermore, we use the same technique to recover the most relevant labels for the clusters. We implemented this approach in a tool called Hapax, which is built on top of the Moose reengineering environment [1]. Figure 1 emphasizes the interactive nature of our tool. On the top side we show the main window of Hapax. On left part of the window is the correlation matrix visualization: the darker the dot, the more similar the two entities. The top two panels on the right show the entities on the current row and column. The bottom-right panel shows the labels attached to the current cluster. On the right side of the window there is a slider for setting the clustering threshold. When the slider is moved, the picture is redrawn with the new clusters. On the bottom side of the figure we show how we use LSI to also search over the entities in the system. The top window contains the search query and the result is shown in the below window with the group of the entities ordered by their relevancy to the query.