Research map / Vision-language and multimodal / Human and embodied modeling
Language Features Matter: Effective Language Representations for Vision-Language Tasks
Andrea Burns, Reuben Tan, Kate Saenko, Stan Sclaroff, Bryan A. Plummer
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use pretrained Word2Vec instead of random scratch embedding (ablation) | LSTM language model · Flickr30K image-sentence retrieval | mean recall: 60.0 -> 68.5 | Improved |
| fine-tune Word2Vec Self-Attention embedding during training (ablation) | Self-Attention language model · Flickr30K image-sentence retrieval | mean recall: 63.6 -> 71.9 | Improved |
| use Self-Attention instead of LSTM on frozen Word2Vec (ablation) | Embedding Network · Flickr30K image-sentence retrieval | mean recall: 68.5 -> 63.6 | No clear effect |
| replace LSTM with Self-Attention language model (fine-tuned Word2Vec) (ablation) | Embedding Network · Flickr30K image-sentence retrieval | mean recall: 69.0 -> 71.9 | Improved |
| use BERT sentence embedding instead of fine-tuned Word2Vec (ablation) | Embedding Network · Flickr30K image-sentence retrieval | mean recall: 71.9 -> 71.8 | Hurt |
| use InferSent instead of BERT for Flickr30K Entities (ablation) | CITE · Flickr30K Entities phrase grounding | accuracy: 69.38 -> 57.83 | Hurt |
| add WordNet and Visual Genome graph retrofitting (GrOVLE) to Word2Vec | Self-Attention + ft · Flickr30K image-sentence retrieval | mean recall: 71.9 -> 72.1 | Improved |
| use Visual Word2Vec instead of GrOVLE (ablation) | Self-Attention + ft · Flickr30K image-sentence retrieval | mean recall: 72.1 -> 68.8 | Hurt |
| add adapted embeddings (GrOVLE) to generation tasks | ARNet captioning · MSCOCO captioning | CIDEr: 94.0 -> 92.5 | No clear effect |
| add multi-task pretraining w/ target task + fine-tuning to GrOVLE (ablation) | Embedding Network · Flickr30K image-sentence retrieval | mean recall: 64.7 -> 72.6 | Improved |
| add multi-task pretraining w/ target task + fine-tuning to GrOVLE (ablation) | Embedding Network · MSCOCO image-sentence retrieval | mean recall: 75.0 -> 81.3 | Improved |
| add multi-task pretraining w/ target task + fine-tuning to GrOVLE (ablation) | CITE · ReferIt phrase grounding | accuracy: 52.15 -> 54.51 | Improved |
Closest papers by meaning
- Revisiting the Role of Language Priors in Vision-Language Models (ICML 2024 · similarity 0.93)
- VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks (ICLR 2025 · similarity 0.91)
- Bridging Vision and Language Spaces with Assignment Prediction (ICLR 2024 · similarity 0.91)
- Open-ended VQA benchmarking of Vision-Language models by exploiting Classification datasets and their semantic hierarchy (ICLR 2024 · similarity 0.90)
- Language-Instructed Vision Embeddings for Controllable and Generalizable Perception (ICLR 2026 · similarity 0.90)
- The Neglected Tails in Vision-Language Models (CVPR 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
