Research map / Image generation / GANs
LocalGAN: Modeling Local Distributions for Adversarial Response Generation
Baoxun Wang, Zhen Xu, Huan Zhang, Kexin Qiu, Deyuan Zhang, Chengjie Sun
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 the DBM free-energy local distribution term instead of Euclidean distance ratio (ablation) | LocalGAN · OpenSubtitles | Dist-1: 0.0314 -> 0.036 | Improved |
| use the DBM free-energy local distribution term instead of Euclidean distance ratio (ablation) | LocalGAN · OpenSubtitles | Ent4: 5.922 -> 6.073 | Improved |
| use the DBM free-energy local distribution term instead of Euclidean distance ratio (ablation) | LocalGAN · Weibo | Ent4: 8.134 -> 8.561 | Improved |
| use the DBM free-energy local distribution term instead of Euclidean distance ratio (ablation) | LocalGAN · Weibo | relevance: 0.311 -> 0.327 | Improved |
| use the DBM free-energy local distribution term instead of Euclidean distance ratio (ablation) | LocalGAN · human evaluation | informativeness: 0.437 -> 0.536 | Improved |
| add the local-distribution hybrid loss to the GAN objective (beta 0 -> 0.9) (ablation) | LocalGAN · OpenSubtitles and Weibo | Dist-1, Dist-2, Ent4, relevance | Improved |
| use only the local-distribution term (beta = 1.0) (ablation) | LocalGAN · OpenSubtitles and Weibo | Dist-1, Dist-2, Ent4, relevance | Hurt |
| add adversarial learning to Seq2Seq (Adver-REGS vs Seq2Seq) | Seq2Seq · OpenSubtitles | Ent4: 5.650 -> 5.701 | Improved |
| add BigGAN operations (larger batch, doubled parameters, orthogonal regularization) to Adver-REGS | Adver-REGS · Weibo | Ent4: 7.658 -> 8.273 | Improved |
| replace baselines with LocalGAN for training stability | adversarial NRG · OpenSubtitles | Dist-1 std: 0.0004 -> 0.0006 | Improved |
| add LocalGAN over strongest baseline BigGAN | BigGAN · OpenSubtitles | relevance: 1.117 -> 1.132 | Improved |
| add LocalGAN over strongest baseline BigGAN | BigGAN · human evaluation | informativeness: 0.487 -> 0.536 | Improved |
Closest papers by meaning
- Posterior-GAN: Towards Informative and Coherent Response Generation with Posterior Generative Adversarial Network (AAAI 2020 · similarity 0.88)
- Generating Informative and Diverse Conversational Responses via Adversarial Information Maximization (NeurIPS 2018 · similarity 0.87)
- Are You Talking to Me? Reasoned Visual Dialog Generation Through Adversarial Learning (CVPR 2018 · similarity 0.87)
- Training Language GANs from Scratch (NeurIPS 2019 · similarity 0.85)
- Meta-CoTGAN: A Meta Cooperative Training Paradigm for Improving Adversarial Text Generation (AAAI 2020 · similarity 0.85)
- Measuring Non-Adversarial Reproduction of Training Data in Large Language Models (ICLR 2025 · similarity 0.84)
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.
