Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CSSG: Measuring Code Similarity with Semantic Graphs

Jingwen Xu, Yiyang Lu, Changze Lv, Zisu Huang, Zhengkang Guo, Zhengyuan Wang, Muzhao Tian, Xuanjing Huang, Xiaoqing Zheng

TL;DR

CSSG addresses the limitation of surface- and syntax-focused code similarity by representing code as integrated program dependence graphs that capture data and control dependencies. It computes semantic similarity via a constrained graph edit distance and normalization, enabling cross-language semantic comparisons. On CodeContests+, CSSG achieves stronger discrimination between correct and incorrect solutions in both monolingual and cross-lingual settings, with stable positive effects across language pairs, highlighting the value of dependency-aware representations. This semantics-grounded approach offers a practical tool for evaluating code similarity and generation quality beyond surface lexical cues.

Abstract

Existing code similarity metrics, such as BLEU, CodeBLEU, and TSED, largely rely on surface-level string overlap or abstract syntax tree structures, and often fail to capture deeper semantic relationships between programs.We propose CSSG (Code Similarity using Semantic Graphs), a novel metric that leverages program dependence graphs to explicitly model control dependencies and variable interactions, providing a semantics-aware representation of code.Experiments on the CodeContests+ dataset show that CSSG consistently outperforms existing metrics in distinguishing more similar code from less similar code under both monolingual and cross-lingual settings, demonstrating that dependency-aware graph representations offer a more effective alternative to surface-level or syntax-based similarity measures.

CSSG: Measuring Code Similarity with Semantic Graphs

TL;DR

CSSG addresses the limitation of surface- and syntax-focused code similarity by representing code as integrated program dependence graphs that capture data and control dependencies. It computes semantic similarity via a constrained graph edit distance and normalization, enabling cross-language semantic comparisons. On CodeContests+, CSSG achieves stronger discrimination between correct and incorrect solutions in both monolingual and cross-lingual settings, with stable positive effects across language pairs, highlighting the value of dependency-aware representations. This semantics-grounded approach offers a practical tool for evaluating code similarity and generation quality beyond surface lexical cues.

Abstract

Existing code similarity metrics, such as BLEU, CodeBLEU, and TSED, largely rely on surface-level string overlap or abstract syntax tree structures, and often fail to capture deeper semantic relationships between programs.We propose CSSG (Code Similarity using Semantic Graphs), a novel metric that leverages program dependence graphs to explicitly model control dependencies and variable interactions, providing a semantics-aware representation of code.Experiments on the CodeContests+ dataset show that CSSG consistently outperforms existing metrics in distinguishing more similar code from less similar code under both monolingual and cross-lingual settings, demonstrating that dependency-aware graph representations offer a more effective alternative to surface-level or syntax-based similarity measures.
Paper Structure (17 sections, 5 equations, 5 figures, 3 tables)

This paper contains 17 sections, 5 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction and Related Work
  2. Method
  3. Sematic Graph Construction
  4. Graph Matching Constraints
  5. Graph Edit Distance
  6. Similarity Normalization
  7. Experiments
  8. General Setups
  9. Dataset
  10. Results
  11. Monolingual Similarity Results
  12. Cross-lingual Similarity Results
  13. Conclusion
  14. Comparison of AST and Semantic Graph Representations
  15. Impact of Data Flow
  16. ...and 2 more sections

Figures (5)

  • Figure 1: Computation pipeline for semantic graph edit distance. The process consists of four main steps: (1) utilizing Joern to extract function-level PDGs from the code snippet; (2) constructing a unified semantic graph by optimizing nodes and integrating a global root with function call edges; (3) computing the constrained Graph Edit Distancebetween the reference and target graphs; and (4) calculating the final similarity score by normalizing the GED against the maximum edit distance.
  • Figure 2: Pearson correlation heatmap among metrics on monolingual code pairs. The correlation patterns suggest that CSSG emphasizes semantic similarities that differ from surface-level metrics while remaining closely related to AST-based approaches.
  • Figure 3: The impact of data flow.
  • Figure 4: The impact of control flow
  • Figure 5: The impact of function call flow