Research map / Architectures and efficiency / Attention and transformers
DiSAN: Directional Self-Attention Network for RNN/CNN-Free Language Understanding
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 multi-dimensional rather than token-wise attention (ablation) | Word-embedding sentence encoder · SNLI test | accuracy: +3.31% | Helped |
| add forward and backward directional self-attention masks (ablation) | DiSAN without directions · SNLI test | accuracy: 0.96% | Helped |
| add forward and backward directional self-attention masks (ablation) | DiSAN without directions · SST fine-grained test | accuracy | Helped |
| replace Bi-LSTM context encoding with DiSA blocks (ablation) | Bi-LSTM with source-to-token self-attention · SNLI test | accuracy: 0.64% | Helped |
| use DiSAN instead of multi-head attention with source-to-token self-attention | Multi-head sentence encoder · SNLI test | accuracy: 1.45% | Helped |
| use DiSAN instead of the strongest earlier sentence encoder | NSE encoder · SNLI test | accuracy: 1.02% | Helped |
| use DiSAN instead of CNN-Tensor | SST fine-grained sentiment classification · SST fine-grained test | accuracy: 0.52% | Helped |
| use DiSAN instead of Dependency Tree-LSTM | Sentence relatedness model · SICK test | MSE: .2532 (.0052) -> .2879 (.0036) | Hurt |
| use DiSAN instead of AdaSent | MPQA opinion polarity classification · MPQA | accuracy: 90.4 (0.7) -> 90.1 (0.4) | Hurt |
| use Adadelta rather than Adam | DiSAN optimization · NLP tasks in the experiments | performance stability | Helped |
Closest papers by meaning
- Bi-Directional Block Self-Attention for Fast and Memory-Efficient Sequence Modeling (ICLR 2018 · similarity 0.88)
- RNN-Based Sequence-Preserved Attention for Dependency Parsing (AAAI 2018 · similarity 0.88)
- Focused Hierarchical RNNs for Conditional Sequence Processing (ICML 2018 · similarity 0.88)
- Multiple Positional Self-Attention Network for Text Classification (AAAI 2020 · similarity 0.87)
- Generalize Sentence Representation with Self-Inference (AAAI 2020 · similarity 0.86)
- RAT: Bridging RNN Efficiency and Attention Accuracy via Chunk-based Sequence Modeling (NeurIPS 2025 · 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.