LLLMWIKI
ArticleMetadataMain page

Question · Updated 2026-07-14

怎样构造既难又不错误的图文检索负例？

#near-cvpr-2025#multimodal#vision-language#evaluation#survey-needed

Open question

怎样构造既难又不错误的图文检索负例？

This question remains open and needs new primary evidence.

Next evidence need

8Candidate sources1Related topics
View evidence paths →

问题

怎样让 hard negative 只改变一个可验证的属性、关系、词序或修改意图，同时避免把多个合理答案错误地标成负例？

为什么还没有解决

ARO 说明组合性 hard negative 能提供训练信号；SugarCrepe 说明 benchmark 正负文本本身可能有分布捷径；CIR/FashionIQ 又存在多个合理目标和 false negative。难例的“难度”和“标签正确性”需要分开审计。

什么证据会有帮助

  • text-only、image-only、词频/长度启发式与视觉模型的对照。
  • 人工一致性、多真值覆盖、负例来源和关系级标签。
  • 加入 hard negative 前后的分层收益，以及对普通 Recall、长尾和域外任务的副作用。
  • 显式负例的 false-positive rate、多个正确答案、query paraphrase 敏感性和多图组合退化；PinPoint 可作为压力测试模板。

相关页面

Metadata
{
  "id": "question-image-text-retrieval-hard-negatives",
  "type": "question",
  "title": "怎样构造既难又不错误的图文检索负例？",
  "status": "open",
  "created": "2026-07-14",
  "updated": "2026-07-14",
  "tags": [
    "near-cvpr-2025",
    "multimodal",
    "vision-language",
    "evaluation",
    "survey-needed"
  ],
  "question_status": "open",
  "priority": "high",
  "related_topics": [
    "topics/image-text-retrieval"
  ],
  "related_entities": [
    "entities/clip-retrieval-benchmarks"
  ],
  "candidate_sources": [
    "sources/2026-07-14-aro",
    "sources/2026-07-14-sugarcrepe",
    "sources/2026-07-14-cirr",
    "sources/2026-07-14-searle-circo",
    "sources/2026-07-14-cprd",
    "sources/2026-07-14-cola",
    "sources/2026-07-14-cir-robustness",
    "sources/2026-07-14-pinpoint"
  ]
}