Algebras FluencyScore at WMT26 Automated Translation Quality Evaluation

Airana Mongush, Dmitrii Pukhov


Abstract
We describe the Algebras FluencyScore submissions to the WMT26 Automated Translation Quality Evaluation shared task (Tasks 1-2). FluencyScore is a reference-free structured LLM judge: five fluency sub-dimensions, a fixed-weight aggregate, and a diagnostic error record (issue type, severity, problem phrase). This system paper adapts that judge rather than replacing it. For Task 1 the diagnostic record is specialized into a multi-error character-span annotator with deterministic phrase-to-offset alignment, merge / whole-segment post-processing, and a conservative add-only overlay. For Task 2 the calibrated scalar is mapped onto the cESA scale with a WMT25-fitted affine adapter and moment matching (aira_mo). Submitted systems use Gemini 3.5 Flash and cover all 21 official pairs plus challenge sets. Official human-cESA rankings were not available at camera-ready time; we report no ranks and no Codabench verification scores.
Anthology ID:
2026.wmt-main.104
Volume:
Proceedings of the Eleventh Conference on Machine Translation
Month:
October
Year:
2026
Address:
Budapest, Hungary
Editors:
Barry Haddow, Tom Kocmi, Philipp Koehn, Christof Monz
Venues:
WMT | WS
SIG:
SIGMT
Publisher:
Association for Computational Linguistics
Note:
Pages:
1711–1715
Language:
URL:
https://aclanthology.org/2026.wmt-main.104/
DOI:
Bibkey:
Cite (ACL):
Airana Mongush and Dmitrii Pukhov. 2026. Algebras FluencyScore at WMT26 Automated Translation Quality Evaluation. In Proceedings of the Eleventh Conference on Machine Translation, pages 1711–1715, Budapest, Hungary. Association for Computational Linguistics.
Cite (Informal):
Algebras FluencyScore at WMT26 Automated Translation Quality Evaluation (Mongush & Pukhov, WMT 2026)
Copy Citation:
PDF:
https://aclanthology.org/2026.wmt-main.104.pdf