Research map / Architectures and efficiency / Attention and transformers
Multi-Scale Self-Attention for Text Classification
Qipeng Guo, Xipeng Qiu, Pengfei Liu, Xiangyang Xue, Zheng Zhang
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 standard self-attention with multi-scale self-attention | Transformer · SST test | accuracy: 50.4 -> 51.9 | Improved |
| Replace standard self-attention with multi-scale self-attention | Transformer · MTL-16 average | accuracy: 82.78 -> 86.34 | Improved |
| Replace standard self-attention with multi-scale self-attention | Transformer · PTB POS, CoNLL2003 NER, CoNLL2012 NER | accuracy / F1: 96.31 -> 97.24; 86.48 -> 89.38; 83.57 -> 85.26 | Improved |
| Replace standard self-attention with multi-scale self-attention | Transformer · SNLI test | accuracy: 82.2 -> 85.9 | Improved |
| Set the scale-distribution factor α to 0.5, favoring smaller scales in shallow layers (ablation) | Multi-Scale Transformer · SNLI test | accuracy: 84.9 -> 85.9 | Improved |
| Set α to -0.5 or -1.0, increasing global bias in shallow layers (ablation) | Multi-Scale Transformer · SNLI test | accuracy: 85.9 -> 84.7; 85.9 -> 84.3 | Hurt |
| Use a single fixed attention scale instead of multiple scales (ablation) | Multi-Scale Transformer · SNLI test | accuracy: 85.9 -> 84.3 (ω=3); 85.9 -> 83.9 (ω=N/16); 85.9 -> 81.7 (ω=N/8); 85.9 -> 80.7 (ω=N/4) | Hurt |
| Add character-level features (ablation) | Multi-Scale Transformer · PTB POS / CoNLL2003 NER / CoNLL2012 NER | accuracy / F1: 97.24 -> 97.54; 89.38 -> 91.33; 85.26 -> 86.77 | Improved |
| Add a CRF layer (ablation) | Multi-Scale Transformer + Char · PTB POS / CoNLL2003 NER / CoNLL2012 NER | accuracy / F1: 97.54 -> 97.66; 91.33 -> 91.59; 86.77 -> 87.80 | Improved |
| Add the classification node and combine it with max-pooling | Multi-Scale Transformer · Text classification and SNLI | accuracy | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DiSAN — DiSAN: Directional Self-Attention Network for RNN/CNN-Free Language Understanding (AAAI 2018) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Learning Multiscale Transformer Models for Sequence Generation (ICML 2022 · similarity 0.87)
- Hierarchical Self-Attention: Generalizing Neural Attention Mechanics to Multi-Scale Problems (NeurIPS 2025 · similarity 0.86)
- The Closeness of In-Context Learning and Weight Shifting for Softmax Regression (NeurIPS 2024 · similarity 0.85)
- Unveiling Induction Heads: Provable Training Dynamics and Feature Learning in Transformers (NeurIPS 2024 · similarity 0.84)
- Selective Attention: Enhancing Transformer through Principled Context Control (NeurIPS 2024 · similarity 0.84)
- Elliptical Attention (NeurIPS 2024 · 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.
