User-Customizable Transpilation of Scripting Languages
summary
A prototype called DuoGlot translates Python to JavaScript by synthesizing syntactic translation rules from user-provided code snippets and composing them automatically using grammar checks plus unit test feedback. The system reaches about 90% accuracy on GeeksForGeeks, translating 10–20% more programs than other translators, uses 142 inferred rules (≈360 lines of human-written snippets), and resolves ambiguous rule choices via an automated retry loop that incurs about a 5× slowdown.
AI-written summary · about
chapters (11)
- 0:00User-customizable transpilation
- 1:55Motivating code translation example
- 3:57Challenges in code migration
- 5:01Research questions and contributions
- 5:21Dolot translator workflow
- 7:52Rule inference technical details
- 9:58Resolving rule composition ambiguities
- 11:23Dolot design goals and explainability
- 12:06Python-to-JavaScript benchmarks
- 14:13Conclusion and future work
- 14:55Q&A
related talks
23mProgram Skeletons for Automated Program Translation
30mIs Polyglot Programming Really a Thing?
10mPerfection Not Required? Human-AI Partnerships in Code Translation
26mAutomating Language Migration at Scale: Tools, Tradeoffs, and the Role of AI
13mKeynote: Not Getting Lost in Translations - Preview Interview with Daniela Engert
13mPublic Static Void
01h 01mProgramming Abstractions for Orchestration of HPC Scientific Computing
47mThe linguist programmer