WELL: Applying Bug Detectors to Bug Localization via Weakly Supervised Learning

Bug localization is a key software development task, where a developer locates the portion of the source code that must be modified based on the bug report. It is label-intensive and time-consuming due to the increasing size and complexity of the modern software. Effectively automating this task can greatly reduce costs by cutting down the developers' effort. Researchers have already made efforts to harness the great powerfulness of deep learning (DL) to automate bug localization. However, training DL models demands a large quantity of annotated training data, while the buggy-location-annotated dataset with reasonable quality and quantity is difficult to collect. This becomes an obstacle to the effective usage of DL for bug localization. We notice that the data pairs for bug detection, which provide weak buggy-or-not binary classification supervision, are much easier to obtain. Inspired by weakly supervised learning, this paper proposes WEakly supervised bug LocaLization (WELL), an approach to transform bug detectors to bug locators. Through the CodeBERT model finetuned by bug detection, WELL is capable to locate bugs in a weakly supervised manner based on the attention. The evaluations on three datasets of WELL show competitive performance with the existing strongly supervised DL solutions. WELL even outperforms current SOTA models in tasks of variable misuse and binary operator misuse.