Similar Scores, Opposite Failures: Human Correction as an Instrument for Video Caption Evaluation
Source check: Source checked. Not independent scientific confirmation.
Finding: Reference-overlap metrics poorly distinguish accepted from rejected captions (BLEU-4 AUC 0.510); window-level review detects more camera/shot errors than overview review. Six models, 196 clips, and human post-edits under overview/window protocols provide the reference; one annotator per caption and this corpus/prompt bound generalization.
Technology: Six frontier video-captioning models generate captions; trained annotators post-edit them against a predefined error taxonomy.
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- Six frontier video-captioning models generate captions; trained annotators post-edit them against a predefined error taxonomy.
- Test and result
- Six models, 196 clips, and human post-edits under overview/window protocols provide the reference; one annotator per caption and this corpus/prompt bound generalization.
- Limitation / still unknown
- The reviewed source does not establish prospective benefit, broad external validity, or replication beyond the reported study.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/html/38511v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review date: 22 Nov 2026. Review scope: Primary result-bearing material for the stated method, result and evaluation, with bibliographic version information..
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.