Research map / Image generation / Text-to-image diffusion
CoNT: Contrastive Neural Text Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Train with CoNT instead of MLE | Transformer-small · WMT'16 Ro-En | BLEU: 25.78 -> 27.70 | Helped |
| Train with CoNT instead of CLAPS | T5-small · WMT'16 Ro-En | BLEU: 29.41 -> 30.91 | Helped |
| Add sequence similarity to decoding (ablation) | CoNT with Transformer-small · IWSLT'14 De-En | BLEU: 34.69 -> 35.55 | Helped |
| Add sequence similarity to decoding (ablation) | CoNT with T5-small · WMT'16 Ro-En | BLEU: 30.54 -> 30.91 | Helped |
| Add sequence similarity to decoding (ablation) | CoNT with Transformer-base · WMT'14 En-De | BLEU: 27.77 -> 28.04 | Helped |
| Use N-Pairs rather than InfoNCE contrastive loss (ablation) | CoNT · IWSLT'14 De-En | BLEU | Helped |
| Use self-generated hypotheses as negative samples rather than only batch samples (ablation) | CoNT · IWSLT'14 De-En | BLEU: +1.0 BLEU | Helped |
| Use batch samples rather than self-generated hypotheses (ablation) | Naive CL · IWSLT'14 De-En | BLEU: less than 0.5 BLEU | Mixed |
| Train with T5-CoNT instead of T5-CLAPS | T5-small · XSum | ROUGE-1: 37.89 -> 39.66 | Helped |
| Apply CoNT to PEGASUS-large | PEGASUS-large · Multi-News | ROUGE-1: 47.52 -> 48.68 | Helped |
| Apply CoNT with T5-base instead of T5-3B | Totto · test set (all) | PARENT: 58.4 -> 58.9 | Helped |
| Apply CoNT with T5-base instead of T5-large | CommonGen · test | CIDEr: 15.13 -> 15.96 | Helped |
Closest papers by meaning
- A Contrastive Framework for Neural Text Generation (NeurIPS 2022 · similarity 0.90)
- Contrastive Learning with Adversarial Perturbations for Conditional Text Generation (ICLR 2021 · similarity 0.85)
- Open-Domain Text Evaluation via Contrastive Distribution Methods (ICML 2024 · similarity 0.84)
- Reflection-Window Decoding: Text Generation with Selective Refinement (ICML 2025 · similarity 0.84)
- Successor Features for Efficient Multi-Subject Controlled Text Generation (ICML 2024 · similarity 0.83)
- Contrastive Learning with Simplicial Convolutional Networks for Short-Text Classification (ICML 2025 · similarity 0.83)
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.