Research map / Architectures and efficiency / Attention and transformers
Segatron: Segment-Aware Transformer for Language Modeling and Understanding
He Bai, Peng Shi, Jimmy Lin, Yuqing Xie, Luchen Tan, Kun Xiong, Wen Gao, Ming Li
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 |
|---|---|---|---|
| Add paragraph and sentence relative position encodings | Transformer-XL base · WikiText-103 | perplexity: 24.4 -> 22.5 | Improved |
| Add paragraph and sentence relative position encodings | Transformer-XL large · WikiText-103 | perplexity: 18.3 -> 17.1 | Improved |
| Add segment-aware position encodings | BERT-base⁻ · GLUE dev | AVG: 82.0 -> 82.8 | Improved |
| Add segment-aware position encodings | BERT-large · GLUE dev | AVG: 85.3 -> 86.4 | Improved |
| Add segment-aware position encodings | BERT-base⁻ · GLUE test | AVG: 77.1 -> 78.7 | Improved |
| Add segment-aware position encodings | BERT-large · GLUE test | AVG: 81.6 -> 82.7 | Improved |
| Use SegaBERT-large for sentence representation learning | S-BERT-large · STS zero-shot tasks | average Spearman's rank correlation ρ×100: 76.68 -> 77.37 | Improved |
| Add segment-aware position encodings | BERT-base⁻ · SQuAD 1.1 | EM: 81.9 -> 83.2 | Improved |
| Add segment-aware position encodings | BERT-base⁻ · SQuAD 2.0 | F1: 78.2 -> 79.2 | Improved |
| Add segment-aware position encodings | BERT-large · SQuAD 1.1 | EM: 86.7 -> 86.0 | Hurt |
| Add segment-aware position encodings | BERT-large · RACE | accuracy: 72.0 -> 73.8 | Improved |
| Add paragraph position encoding only (ablation) | Transformer-XL base · WikiText-103 | perplexity: 24.35 -> 24.07 | Improved |
Closest papers by meaning
- Distilling Sequential Computation in Transformer Language Models (NeurIPS 2026 · similarity 0.85)
- LeaPformer: Enabling Linear Transformers for Autoregressive and Simultaneous Tasks via Learned Proportions (ICML 2024 · similarity 0.85)
- Transformer Block Coupling and its Correlation with Generalization in LLMs (ICLR 2025 · similarity 0.84)
- Beyond Attention: Breaking the Limits of Transformer Context Length with Recurrent Memory (AAAI 2024 · similarity 0.84)
- Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models (ICLR 2025 · similarity 0.84)
- What Rotary Position Embedding Can Tell Us: Identifying Query and Key Weights Corresponding to Basic Syntactic or High-level Semantic Information (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.
