Research map / Image generation / Autoregressive and latent generation
Educating Text Autoencoders: Latent Representation Guidance via Denoising
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 the AAE with word-deletion denoising on inputs | text autoencoders · Yelp and Yahoo neighborhood preservation | nearest-neighbor recall: about 1.5~2 times that of AAE | Helped |
| Use input word deletion rather than standard AAE training | AAE · Yelp generation-reconstruction trade-off | BLEU and PPL trade-off | Helped |
| Use DAAE instead of β-VAE | Yelp tense transfer · Yelp tense transfer | ACC: 49.0 -> 50.3 | Helped |
| Use DAAE instead of AAE | Yelp tense transfer · Yelp tense transfer | PPL: 37.4 -> 32.0 | Helped |
| Use DAAE instead of AAE | Yelp tense transfer · Yelp tense transfer | BLEU: 82.2 -> 54.3 | Hurt |
| Scale the DAAE sentiment vector from ±v to ±1.5v (ablation) | DAAE · Yelp sentiment transfer | ACC: 36.2 -> 73.6 | Helped |
| Scale the DAAE sentiment vector from ±v to ±1.5v (ablation) | DAAE · Yelp sentiment transfer | PPL: 40.0 -> 54.1 | Hurt |
| Scale the DAAE sentiment vector from ±v to ±2v (ablation) | DAAE · Yelp sentiment transfer | ACC: 36.2 -> 91.8 | Helped |
| Scale the DAAE sentiment vector from ±v to ±2v (ablation) | DAAE · Yelp sentiment transfer | BLEU: 40.9 -> 7.3 | Hurt |
| Scale the DAAE sentiment vector from ±v to ±2v (ablation) | DAAE · Yelp sentiment transfer | PPL: 40.0 -> 61.8 | Hurt |
| Scale the AAE sentiment vector from ±v to ±2v (ablation) | AAE · Yelp sentiment transfer | ACC: 7.2 -> 57.5 | Helped |
| Scale the AAE sentiment vector from ±v to ±2v (ablation) | AAE · Yelp sentiment transfer | PPL: 33.7 -> 139.8 | Hurt |
Closest papers by meaning
- Adversarially Regularized Autoencoders (ICML 2018 · similarity 0.90)
- Toward Controlled Generation of Text (ICML 2017 · similarity 0.89)
- Towards Text Generation with Adversarially Learned Neural Outlines (NeurIPS 2018 · similarity 0.88)
- Compressed and Smooth Latent Space for Text Diffusion Modeling (NeurIPS 2025 · similarity 0.86)
- How to Train Your Latent Diffusion Language Model Jointly With the Latent Space (NeurIPS 2026 · similarity 0.86)
- DenoSent: A Denoising Objective for Self-Supervised Sentence Representation Learning (AAAI 2024 · similarity 0.86)
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.