Dependency-aware Code Naturalness
summary
Proposes DAN that builds a program dependency graph, extracts DFS paths then fixed-length subpaths, and measures sequence-level naturalness with n‑gram and CodeBERT models. Evaluated on HumanEvalX with three synthetic degradations (dead‑code insertion, confusing insertion, variable renaming), DAN’s n‑gram measure shows about a 41% average improvement over per-line n‑grams and about 58% gain versus a whole-file CSN_sub baseline; it also raises bug‑line prioritization and yields cleaner training data that modestly improves CodeBLEU and pass rates.
AI-written summary · about
chapters (7)
- 0:00Code naturalness research
- 4:54Limitations of models
- 5:56DepGram for dependency-aware measurement
- 8:26DepGram's effectiveness
- 12:44DepGram on downstream tasks
- 17:30Conclusion and future
- 19:15Q&A
related talks
47mExploring a complex codebase with AI
16mDemystifying Dependence
41mDomain Re discovery Patterns for Legacy Code
26mDetecting and Mapping Code Defects Introduced by Generative AI Using CodeLogic & Neo4j
15mAutomatic Synthesis of Data-Flow Analyzers
14mNeural Reverse Engineering of Stripped Binaries using Augmented Control Flow Graphs
15mAdversarial Examples for Models of Code
12mEnhancing Code Vulnerability Detection via Vulnerability-Preserving Data Augmentation