Researchers revisit round-trip translation for reference-free machine translation evaluation
An arXiv paper argues that round-trip translation, long viewed as unreliable for evaluation, can work without reference translations once a copying mechanism is removed.
In plain words
Checking a translation usually requires a human-made reference answer, which is scarce for low-resource languages. The authors say round-trip translation—translating text to another language and back—can score translations without references. They trace earlier doubts to a copying mechanism in older systems.
What was announced
The paper, published on arXiv, reports that round-trip translation can serve automatic evaluation without references. The authors revisit why forward and round-trip scores correlated poorly in the statistical machine translation era and attribute the long-standing misunderstanding to the copying mechanism. After removing that mechanism, round-trip scores reflected forward translation performance appropriately, they report.
The authors describe the correction as overdue and outline three uses: predicting forward translation scores, improving recently advanced quality estimation models, and identifying adversarial competitors in shared tasks through cross-system verification. The work targets low-resource settings, where a lack of parallel corpora limits conventional evaluation. The claims rest on the authors' own experiments described in the abstract; the supplied material does not include the full paper's results, datasets or independent replication.
What it does not show
- Evidence is limited to the paper's abstract; experimental details, datasets and full results are not available in the supplied passages.
- The findings are the authors' own and are not independently confirmed.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.