Research map / Architectures and efficiency / Attention and transformers
Bi-Directional Block Self-Attention for Fast and Memory-Efficient Sequence Modeling
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 block self-attention with intra-block and inter-block context | Bi-BloSAN · SNLI | accuracy: 85.0 -> 85.7 | Helped |
| Use block self-attention instead of DiSAN | sequence encoding on SNLI · SNLI | training GPU memory: 2267 -> 1243 | Helped |
| Use block self-attention instead of Bi-LSTM | modified SQuAD sentence locating · Modified SQuAD dev | accuracy: 68.01 -> 68.38 | Helped |
| Use Bi-BloSAN rather than DiSAN | SICK sentence relatedness · SICK | Pearson's r: 0.8695 (0.0012) -> 0.8616 (0.0012) | Hurt |
| Use Bi-BloSAN rather than Bi-LSTM | sentence classification · SUBJ | accuracy: 94.7 (0.7) -> 94.5 (0.5) | Hurt |
| Use Bi-BloSAN rather than DiSAN | sentence classification · SST-2 | accuracy: 51.0 (0.7) -> 50.6 (0.5) | Hurt |
| Add local context representations (ablation) | Bi-BloSAN · SNLI | accuracy: 85.2 -> 85.7 | Helped |
| Add global context representations (ablation) | Bi-BloSAN · SNLI | accuracy: 85.3 -> 85.7 | Helped |
| Add mBloSA context fusion (ablation) | Bi-BloSAN · SNLI | accuracy: 83.1 -> 85.7 | Helped |
| Add source2token self-attention instead of vanilla attention (ablation) | Bi-BloSAN without mBloSA · SNLI | accuracy: +3.3% | Helped |
Closest papers by meaning
- DiSAN: Directional Self-Attention Network for RNN/CNN-Free Language Understanding (AAAI 2018 · similarity 0.88)
- RAT: Bridging RNN Efficiency and Attention Accuracy via Chunk-based Sequence Modeling (NeurIPS 2025 · similarity 0.87)
- Focused Hierarchical RNNs for Conditional Sequence Processing (ICML 2018 · similarity 0.87)
- Word Attention for Sequence to Sequence Text Understanding (AAAI 2018 · similarity 0.87)
- Improving Bilinear RNN with Closed-loop Control (NeurIPS 2025 · similarity 0.86)
- Explaining Modern Gated-Linear RNNs via a Unified Implicit Attention Formulation (ICLR 2025 · similarity 0.85)
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.