Lingxiao Tang, He Ye, Zhaoyang Chu, Muyang Ye, Zhongxin Liu, Xiaoxue Ren, Lingfeng Bao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The SZZ algorithm and its variants have been extensively utilized for identifying bug-inducing commits based on bug-fixing commits. However, these algorithms face challenges when there are no deletion lines in the bug-fixing commit. Previous studies have attempted to address this issue by tracing back all lines in the block that encapsulates the added lines. However, this method is too coarse-grained and suffers from low precision. To address this issue, we propose a novel method in this paper calledSem-SZZ, which is based on fine-grained semantic analysis. Initially, we observe that a significant number of bug-inducing commits can be identified by tracing back the unmodified lines near added lines, resulting in improved precision and F1-score. Building on this observation, we conduct a more fine-grained semantic analysis. We begin by performing program slicing to extract the program part near the added lines. Subsequently, we compare the program's states between the previous version and the current version, focusing on data flow and control flow differences based on the extracted program part. Finally, we extract statements contributing to the bug based on these differences and utilize them to locate bug-inducing commits. We also extend our approach to fit the scenario where the bug-fixing commits contain deleted lines. Experimental results demonstrate thatSem-SZZoutperforms the state-of-the-art methods in identifying bug-inducing commits, regardless of whether the bug-fixing commit contains deleted lines.

The SZZ algorithm has been widely used for identifying bug-inducing commits. However, it suffers from low precision, as not all deletion lines in the bug-fixing commit are related to the bug fix. Previous studies have attempted to address this issue by using static methods to filter out noise, e.g., comments and refactoring operations in the bug-fixing commit. However, these methods have two limitations. First, it is challenging to include all refactoring and non-essential change patterns in a tool, leading to the potential exclusion of relevant lines and the inclusion of irrelevant lines. Second, applying these tools might not always improve performance. In this paper, to address the aforementioned challenges, we propose NEURALSZZ, a deep learning approach for detecting the root cause deletion lines in a bug-fixing commit and using them as input for the SZZ algorithm. NEURALSZZ first constructs a heterogeneous graph attention network model that captures the semantic relationships between each deletion line and the other deletion and addition lines. To pinpoint the root cause of a bug, Neuralszz uses a learning-to-rank technique to rank all deletion lines in the commit. To evaluate the effectiveness of NEURALSZZ, we utilize three datasets containing high-quality bug-fixing and bug-inducing commits. The experiment results show that NEURALSZZ outperforms various baseline methods, e.g., traditional machine learning-based approaches and BI-LSTM in identifying the root cause of bugs. Moreover, by utilizing the top-ranked deletion lines and applying the SZZ algorithm, Neuralszz demonstrates better precision and F1-score compared to previous SZZ algorithms.