ML-Enhanced Code Completion Improves Developer Productivity (2022)
a year ago
- #developer tools
- #code completion
- #machine learning
- Google developed a hybrid semantic ML code completion tool combining machine learning and rule-based semantic engines.
- The tool uses Transformer models trained on Google's monorepo across eight programming languages, improving developer productivity.
- Single-line ML completion reduced coding iteration time by 6% among 10k+ Google developers.
- ML-enhanced suggestions now account for 3% of new code generated at Google.
- Semantic correctness checks improved the acceptance rate of ML suggestions by filtering out uncompilable code.
- The integration of ML and semantic engines allows for both single and multi-line code completions.
- Future work includes further collaboration between ML models and semantic engines for long predictions and API exploration.