A deep multimodal model for bug localization

摘要

Bug localization utilizes the collected bug reports to locate the buggy source files. The state of the art falls short in handling the following three aspects, including (L1) the subtle difference between natural language and programming language, (L2) the noise in the bug reports and (L3) the multi-grained nature of programming language. To overcome these limitations, we propose a novel deep multimodal model named DeMoB for bug localization. It embraces three key features, each of which is tailored to address each of the three limitations. To be specific, the proposed DeMoB generates the multimodal coordinated representations for both bug reports and source files for addressing L1. It further incorporates the AttL encoder to process bug reports for addressing L2, and the MDCL encoder to process source files for addressing L3. Extensive experiments on four large-scale real-world data sets demonstrate that the proposed DeMoB significantly outperforms existing techniques.