Lune

ACL Findings2023

On the Difference of BERT-style and CLIP-style Text Encoders

Zhihong Chen, Guiming Hardy Chen, Shizhe Diao, Xiang Wan, Benyou Wang

摘要

Masked language modeling (MLM) has been one of the most popular pretraining recipes in natural language processing, e.g., BERT, one of the representative models. Recently, contrastive language-image pretraining (CLIP) has also attracted attention, especially its vision models that achieve excellent performance on a broad range of vision tasks. However, few studies are dedicated to studying the text encoders learned by CLIP. In this paper, we analyze the difference between BERT-style and CLIP-style text encoders from three experiments: (i) general text understanding, (ii) vision-centric text understanding, and (iii) text-to-image generation. Experimental analyses show that although CLIP-style text encoders underperform BERTstyle ones for general text understanding tasks, they are equipped with a unique ability, i.e., synesthesia, for the cross-modal association, which is more similar to the senses of humans. Our code is released at https://github.com/ zhjohnchan/probing-clip-dev.

问 Lune

几分钟读透这篇 Paper,不用再花一下午。

Lune 能读完全文,也能读引用它的顶会 Paper,讲清它的结论、证据和局限,每一点都注明出自哪篇 Paper。不用硬啃,也不靠猜。

已有 2 篇顶会 Paper 引用了它。别等审稿人问起,先弄清它们发现了什么。

可以从这些问题问起

Lune 可以查阅

全文 · 引用它的顶会 Paper · 相关的 Paper

获取答案

免费开始,无需绑卡

脉络

  • Visual-Text Cross Alignment: Refining the Similarity Score in Vision-Language Models 引用了 Learning Transferable Visual Models From Natural Language Supervision
  • Decoder-Only LLMs are Better Controllers for Diffusion Models 引用了 Learning Transferable Visual Models From Natural Language Supervision
  • When and Why Vision-Language Models Behave like Bags-Of-Words, and What to Do About It? 引用了 Learning Transferable Visual Models From Natural Language Supervision
  • Learning Transferable Visual Models From Natural Language Supervision 引用了 An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
它引用的引用它的从左到右按年份排列,引用越多,位置越高。

引用它的顶会 Paper2

  1. Visual-Text Cross Alignment: Refining the Similarity Score in Vision-Language ModelsJinhao Li, Haopeng Li, Sarah Monazam Erfani, Lei Feng 等ICML 2024 · 30 次引用
  2. Decoder-Only LLMs are Better Controllers for Diffusion ModelsZiyi Dong, Yao Xiao, Pengxu Wei, Liang LinACM MM 2024 · 3 次引用

它引用的 Paper6

  1. Learning Transferable Visual Models From Natural Language SupervisionAlec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh 等ICML 2021 · 47,906 次引用
  2. An Image is Worth 16x16 Words: Transformers for Image Recognition at ScaleAlexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn 等ICLR 2021 · 21,477 次引用
  3. ALBERT: A Lite BERT for Self-supervised Learning of Language RepresentationsZhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel 等ICLR 2020 · 7,418 次引用
  4. When and Why Vision-Language Models Behave like Bags-Of-Words, and What to Do About It?Mert Yüksekgönül, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky 等ICLR 2023 · 37 次引用
  5. TediGAN: Text-Guided Diverse Face Image Generation and ManipulationWeihao Xia, Yujiu Yang, Jing-Hao Xue, Baoyuan WuCVPR 2021

相关 Paper

  1. CLIP Behaves like a Bag-of-Words Model Cross-modally but not Uni-modallyDarina Koishigarina, Arnas Uselis, Seong Joon OhICLR 2026 · 33 次引用
  2. On the Language Encoder of Contrastive Cross-modal ModelsMengjie Zhao, Junya Ono, Zhi Zhong, Chieh-Hsin Lai 等ACL Findings 2024
  3. Understanding Transferable Representation Learning and Zero-shot Transfer in CLIPZixiang Chen, Yihe Deng, Yuanzhi Li, Quanquan GuICLR 2024 · 21 次引用
  4. un2CLIP: Improving CLIP's Visual Detail Capturing Ability via Inverting unCLIPYinqi Li, Jiahe Zhao, Hong Chang, Ruibing Hou 等NeurIPS 2025 · 6 次引用
  5. MaskCLIP: Masked Self-Distillation Advances Contrastive Language-Image PretrainingXiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang 等CVPR 2023
黄昏的海面,两侧是细线勾勒的悬崖