Classifying Bugs with Interpolants

We present an approach to the classification of error messages in the context of static checking in the style of ESC/Java. The idea is to compute a semantics-based signature for each error message and then group together error messages with the same signature. The approach aims at exploiting modern verification techniques based on, e.g., Craig interpolation in order to generate small but significant signatures. We have implemented the approach and applied it to three benchmark sets (from Apache Ant, Apache Cassandra, and our own tool). Our experiments indicate an interesting practical potential. More than half of the considered error messages (for procedures with more than just one error message) can be grouped together with another error message.