We presented the findings and overview of the SIGTURK 2026 Shared Task at the SIGTURK Workshop during EACL 2026. The shared task challenged teams on three subtasks: detecting and aligning technical terms, correcting term translations using expert-provided hints, and full end-to-end post-editing with an offline terminology glossary. A fine-tuned 4B model beat our best baseline (GPT-5.2) on end-to-end post-editing, but models still fall well behind human experts when it comes to following expert terminology.
- Leaderboard
- Repo
- Our paper is available as a PDF here.