TransMap: Pinpointing Mistakes in Neural Code Translation

• DOI: 10.1145/3611643.3616322
• Publication Date: 2023-11-30T23:14:38Z
• Authors: Bo Wang; Ruishi Li; Mingkai Li; Prateek Saxena
• Keywords: Python; Mistake; Code (set theory)

Automated code translation between programming languages can greatly reduce the human effort needed in learning new or migrating code. Recent neural machine models, such as Codex, have been shown to be effective on many generation tasks including translation. However, produced by translators often has semantic mistakes. These mistakes are difficult eliminate from translator itself because is a black box, which interpret control compared rule-based transpilers. We propose first automated approach pinpoint obtained after Our techniques implemented prototype tool called TransMap translates Python JavaScript, both of popular scripting languages. On our created micro-benchmarks programs with 648 total, accurately pinpoints correct location for fix 87.96%, highlighting 1-2 lines user inspect per mistake. report experience translating 5 libraries up 1k TransMap. preliminary study suggests that time fixing around 70% using standard IDE debuggers.