Research map / Image generation / Autoregressive and latent generation
Towards Text Generation with Adversarially Learned Neural Outlines
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 |
|---|---|---|---|
| increase Gaussian noise added to sentence vectors when training decoder (ablation) | GRU decoder · SNLI | reconstruction BLEU-4: 64.54 -> 37.51 | Hurt |
| increase decoder training noise (ablation) | GRU decoder · SNLI | sample quality | Helped |
| use generated outlines instead of WD-LSTM at temperature 1.0 | outline-conditioned decoder · WMT15 / BookCorpus | human overall preference: 63.95% vs 36.05% (WMT); 70.35% vs 29.65% (BookCorpus) | Helped |
| beam search outline model vs WD-LSTM at temperature 0.5 | outline-conditioned decoder · WMT15 / BookCorpus | human overall preference: 53.20% vs 46.80% (WMT); 57.28% vs 42.72% (BookCorpus) | No clear effect |
| use outline-based conditional GAN vs Seq2Seq baseline | conditional GAN · SNLI premise generation | NLI accuracy: 62.6% -> 70.8% | Helped |
| use outline GAN vs Shen et al. MOSM | conditional GAN · SNLI premise generation | NLI accuracy: 75.9% -> 70.8% | Hurt |
| add MSE term to adversarial objective (ablation) | conditional GAN · SNLI | quality | No clear effect |
| gradient-based interpolation instead of linear interpolation in encoder space (ablation) | decoder-guided interpolation · BookCorpus | interpolation naturalness | Helped |
Closest papers by meaning
- MaskGAN: Better Text Generation via Filling in the _______ (ICLR 2018 · similarity 0.91)
- Branch-GAN: Improving Text Generation with (not so) Large Language Models (ICLR 2024 · similarity 0.89)
- Adversarial Feature Matching for Text Generation (ICML 2017 · similarity 0.89)
- Long Text Generation via Adversarial Training with Leaked Information (AAAI 2018 · similarity 0.89)
- Educating Text Autoencoders: Latent Representation Guidance via Denoising (ICML 2020 · similarity 0.88)
- Differentiated Distribution Recovery for Neural Text Generation (AAAI 2019 · similarity 0.87)
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.