Skip to content
conftrace_
2026 EMNLP EMNLP 2026

Full-Video Context and Anonymous Candidate Selection for WMT26 Video Subtitle Translation

AbstractWe present ReopenAI’s constrained-track system for the WMT26 Video Subtitle Translation Shared Task. The system translates Simplified Chinese subtitles into English, Thai, Indonesian, Malay, and Traditional Chinese for Taiwan. It combines Hy-MT2-7B and Gemma-4-12B-it without fine-tuning, using approximately 19B unique parameters. Both models receive the complete Chinese subtitle sequence and video titles as context and generate two candidates for each subtitle cue. A Gemma-based selector anonymously compares four candidates in a stable order and its reverse. A selection is accepted only when both judgments agree; otherwise, a deterministic Gemma candidate is used as fallback. In the official preliminary evaluation, ReopenAI ranked first overall and in every target direction, achieving a macro-average score of 89.357 across 100 videos. On an internal three-video diagnostic covering 705 cues per target, the fusion system achieved a 93.8% pooled acceptable-translation rate. A separate 122-cue ablation showed that full-video context improved Hy-MT2 in all five target languages, with gains ranging from 4.1 to 18.0 points. The system uses no task-specific parallel data, audio, or video frames.

More from EMNLP 2026

All EMNLP 2026 papers →