Research map / Architectures and efficiency / Attention and transformers
Generating Sentences Using a Dynamic Canvas
Harshil Shah, Bowen Zheng, David Barber
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 |
|---|---|---|---|
| replace the sequential Gen-RNN generator with AUTR's attention-based dynamic canvas | sentence generation · Book Corpus test | ELBO: -52.3 -> -50.7 | Improved |
| replace the sequential Gen-RNN generator with AUTR's attention-based dynamic canvas | sentence generation · Book Corpus test | PPL: 41.9 -> 37.4 | Improved |
| replace the sequential Gen-RNN generator with AUTR's attention-based dynamic canvas | sentence generation · Book Corpus test | KL: 7.1 -> 8.0 | Improved |
| add 30% word dropout (ablation) | AUTR, T=30 · Book Corpus test | ELBO: -50.7 -> -51.6 | Hurt |
| add 30% word dropout (ablation) | AUTR, T=40 · Book Corpus test | ELBO: -50.7 -> -51.5 | Hurt |
| use T=40 instead of T=30 (ablation) | AUTR without dropout · Book Corpus test | PPL: 37.4 -> 37.4 | No clear effect |
| use T=40 instead of T=30 (ablation) | AUTR with 30% dropout · Book Corpus test | ELBO: -51.6 -> -51.5 | No clear effect |
| add 30% word dropout (ablation) | AUTR · Book Corpus test | KL: 8.0 -> 14.0 | Improved |
| use AUTR instead of Gen-RNN for iterative missing-word imputation | missing-word imputation · Book Corpus test, 30% words missing | missing-word imputation accuracy: 31.9% -> 34.1% | Improved |
| use modified attention that favors unwritten slots and caps cumulative attention (ablation) | AUTR canvas updates · Book Corpus | model performance | Improved |
| use dynamic attention instead of a fixed canvas update mechanism (ablation) | AUTR canvas updates · Book Corpus | use of recurrent computation steps | Improved |
| condition each generated word on preceding words rather than generate words independently (ablation) | AUTR text decoder · Book Corpus sentence generation | local word consistency | Improved |
Closest papers by meaning
- Towards Text Generation with Adversarially Learned Neural Outlines (NeurIPS 2018 · similarity 0.86)
- Toward Controlled Generation of Text (ICML 2017 · similarity 0.85)
- Can Unconditional Language Models Recover Arbitrary Sentences? (NeurIPS 2019 · similarity 0.85)
- Beyond Attention: Breaking the Limits of Transformer Context Length with Recurrent Memory (AAAI 2024 · similarity 0.84)
- Parameterized Synthetic Text Generation with SimpleStories (NeurIPS 2025 · similarity 0.84)
- Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable 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.
