SOFTWARE BUG LOCALIZATION USING TOPIC MODELS

Abstract

Bug localization is a process of identifying the specific file of source code that is faulty and needs to be modified to fix the bug. Due to the increasing size and complexity of current software applications, automated solutions for bug localization can significantly reduce human effort and software development/maintenance cost. In this research work, bug localization has been performed using topic model of Information Retrieval. Pachinko Allocation Model (PAM) has been applied for the first time in bug localization. In this research work, PAM model of source code is built first. This model is then queried for locating bugs. The bug reports are considered as a query for the system for which files containing bugs need to be identified. This query is used by Inference engine to produce ranked list of files from source code. The top-ranked files are the one most likely to require modification to correct the bug. This work performs analysis and comparison of PAM and Latent Dirichlet Allocation (LDA) models based approach for bug localization using MALLET library in java. This library has been extended to incorporate PAM based bug localization using proposed Inference engine. For evaluating the performance of PAM and LDA based approach, the datasets downloaded from two open source projects i.e. Rhino and ModeShape have been used in this work. In case of Rhino dataset, for one bug report only 10% of dataset is needed to be reviewed. In case of ModeShape dataset, for one bug report only 1.5 % of dataset is needed to be reviewed. It has been observed that the bug localization technique using PAM model gives promising results as compared to LDA model.