Research map / Image generation / GANs
Self-Adversarial Learning with Comparative Discrimination for 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 |
|---|---|---|---|
| Use self-play comparisons to previously generated samples instead of comparisons with real samples (ablation) | SAL · Synthetic (sequence length 20) | NLL: 14.65 $\pm{0.16}$ -> 14.29 $\pm{0.11}$ | Helped |
| Use self-play comparisons to previously generated samples instead of comparisons with real samples (ablation) | SAL · COCO | Perplexity: 276.7$\pm{12.5}$ -> 231.3$\pm{10.8}$ | Helped |
| Replace binary discrimination scores with a comparative discriminator (ablation) | SAL · Synthetic (sequence length 20) | NLL: 15.01 $\pm{0.04}$ -> 14.29 $\pm{0.11}$ | Helped |
| Replace binary discrimination scores with a comparative discriminator (ablation) | SAL · COCO | Perplexity: 341.8 $\pm{13.4}$ -> 231.3$\pm{10.8}$ | Helped |
| Add an indistinguishable class to the comparative discriminator (ablation) | SAL · Synthetic (sequence length 20) | NLL: 14.85 $\pm{0.16}$ -> 14.29 $\pm{0.11}$ | Helped |
| Add an indistinguishable class to the comparative discriminator (ablation) | SAL · COCO | Perplexity: 291.5 $\pm{16.3}$ -> 231.3$\pm{10.8}$ | Helped |
| Use scheduled rewarding (ablation) | SAL · Synthetic (sequence length 20) | NLL: 14.46$\pm{0.18}$ -> 14.29 $\pm{0.11}$ | Helped |
| Use scheduled rewarding (ablation) | SAL · COCO | Perplexity: 248.3 $\pm{14.2}$ -> 231.3$\pm{10.8}$ | Helped |
| Use memory replay (ablation) | SAL · Synthetic (sequence length 20) | NLL: 14.41 $\pm{0.17}$ -> 14.29 $\pm{0.11}$ | Helped |
| Use memory replay (ablation) | SAL · COCO | Perplexity: 254.6 $\pm{13.7}$ -> 231.3$\pm{10.8}$ | Helped |
| Use SAL rather than MLE | Text generation · COCO | BLEU-4(F): 0.303 $\pm{0.01}$ -> 0.362 $\pm{0.02}$ | Helped |
| Use SAL rather than MLE | Text generation · COCO | NLL_gen: 0.721 $\pm{0.02}$ -> 0.873 $\pm{0.02}$ | Hurt |
Closest papers by meaning
- Adversarial Feature Matching for Text Generation (ICML 2017 · similarity 0.90)
- TextGAIL: Generative Adversarial Imitation Learning for Text Generation (AAAI 2021 · similarity 0.90)
- MaskGAN: Better Text Generation via Filling in the _______ (ICLR 2018 · similarity 0.89)
- Branch-GAN: Improving Text Generation with (not so) Large Language Models (ICLR 2024 · similarity 0.89)
- Meta-CoTGAN: A Meta Cooperative Training Paradigm for Improving Adversarial Text Generation (AAAI 2020 · similarity 0.89)
- Adversarial Text Generation via Feature-Mover's Distance (NeurIPS 2018 · similarity 0.89)
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.