Model Editing for Targeted Translation Error Repair
Abstract
AbstractMachine translation systems still make systematic errors on idioms and other expressions that cannot be translated literally. Model editing offers a possible way to fix such recurring errors without retraining the whole model, but it has mostly been studied for monolingual factual updates rather than translation. We introduce MEMT-bench, the first benchmark for evaluating model editing in machine translation across multiple languages, models, and editing settings. The benchmark is built from idiomatic translation errors and tests whether an edit can correct a targeted mistranslation while preserving unrelated translation behavior. Experiments with causal tracing and representative editing methods reveal both the promise and difficulty of translation editing: edits can repair targeted errors, but maintaining cross-lingual locality and robustness under repeated edits remains challenging. Code and data are publicly available at [GitHub](https://github.com/dannalily/MEMT-Model-Editing-for-Targeted-Translation-Error-Repair/) and [Hugging Face](https://huggingface.co/datasets/Dannalily/MEMT-bench), respectively.