A Syntax-Guided Edit Decoder for Neural Program Repair (2106.08253v6)

Abstract: Automated Program Repair (APR) helps improve the efficiency of software development and maintenance. Recent APR techniques use deep learning, particularly the encoder-decoder architecture, to generate patches. Though existing DL-based APR approaches have proposed different encoder architectures, the decoder remains to be the standard one, which generates a sequence of tokens one by one to replace the faulty statement. This decoder has multiple limitations: 1) allowing to generate syntactically incorrect programs, 2) inefficiently representing small edits, and 3) not being able to generate project-specific identifiers. In this paper, we propose Recoder, a syntax-guided edit decoder with placeholder generation. Recoder is novel in multiple aspects: 1) Recoder generates edits rather than modified code, allowing efficient representation of small edits; 2) Recoder is syntax-guided, with the novel provider/decider architecture to ensure the syntactic correctness of the patched program and accurate generation; 3) Recoder generates placeholders that could be instantiated as project-specific identifiers later. We conduct experiments to evaluate Recoder on 395 bugs from Defects4J v1.2 and 420 additional bugs from Defects4J v2.0. Our results show that Recoder repairs 53 bugs on Defects4J v1.2, which achieves 21.4% improvement over the previous state-of-the-art approach for single-hunk bugs (TBar). Importantly, to our knowledge, Recoder is the first DL-based APR approach that has outperformed the traditional APR approaches on this dataset. Furthermore, Recoder also repairs 19 bugs on the additional bugs from Defects4J v2.0, which is 137.5% more than TBar (8 bugs) and 850% more than SimFix (2 bugs). This result suggests that Recoder has better generalizability than existing APR approaches.

• The paper introduces Recoder, an edit-based decoder that generates efficient, syntactically correct patches for automated program repair.
• It employs a provider/decider architecture to guide syntax-aware edit generation, greatly reducing the production of invalid patches.
• Experimental results on benchmarks like Defects4J, IntroClassJava, and QuixBugs demonstrate significant improvements over traditional token-based APR methods.

Syntax-Guided Edit Decoder for Neural Program Repair

The paper "A Syntax-Guided Edit Decoder for Neural Program Repair" presents a novel approach to Automated Program Repair (APR) using deep learning techniques. The primary focus is on enhancing the decoder component within the encoder-decoder architecture, which is pivotal in generating patches for software repair. In existing methodologies, the decoder typically generates a sequence of tokens to replace faulty code with modified statements. However, this paper introduces Recoder—a syntax-guided edit decoder that offers improvements in several facets of program repair.

Key Innovations and Methodology

Recoder presents advancements by departing from the conventional token-sequence approach. The innovations introduced by Recoder are noteworthy:

1. Edit-based Representation: Unlike traditional methods that produce modified code directly, Recoder generates edits. This facilitates a more efficient representation of small code changes, thereby reducing the patch space and aiding in the generation of syntactically correct patches.
2. Syntax-guided Provider/Decider Architecture: This architecture ensures the syntactic correctness of patches and helps in accurate generation. The decider allocates a probability distribution over providers which then generate syntax-guided edits. This approach significantly diminishes the generation of syntactically invalid patches, a common limitation in previous APR methods.
3. Placeholder Generation: Recoder introduces placeholders for identifiers which can be instantiated with project-specific names. This is crucial for generating accurate patches when dealing with unique project-specific symbols, a task traditional methods struggle with.

Experimental Results

The novel approach is empirically evaluated on several benchmarks, demonstrating its effectiveness:

• Defects4J v1.2: Recoder successfully repaired 51 bugs and notably achieved 21.4% improvement (correctly repairing 9 more bugs) compared with the previous state-of-the-art APR approach for single-hunk bugs, TBar.
• Defects4J v2.0: Recoder repaired 19 bugs—137.5% more than TBar and 850% more than SimFix, indicating remarkable generalizability and effectiveness across newer bug benchmarks.
• IntroClassJava and QuixBugs: Recoder showcased its dominance by repairing 775% more bugs in IntroClassJava and 30.8% more bugs in QuixBugs, underscoring its robustness and adaptability across diverse datasets.

Implications and Future Work

The better performance and generalizability suggest Recoder's potential in broadening the reach and applicability of APR in various software projects. This improvement aligns with the ambition of leveraging deep learning for more effective software maintenance solutions. Additionally, it emphasizes the power of syntactic guidance in neural architectures, which could inspire further exploration in neural network-based code synthesis and transformation.

Future work might include extending Recoder for multi-hunk patches, thereby improving its applicability in complex bug scenarios. Also, the exploration of alternative neural architectures and deeper integration of project-specific context could augment Recoder's accuracy and efficiency. This research serves as a foundation for advancing neural-based coding tools, potentially influencing the landscape of intelligent code editing and automatic correction platforms.

In summary, Recoder represents a significant step in APR techniques, offering new pathways in the development of intelligent, syntax-aware neural models that promise to enhance the process of automated code repair and transformation.