Research map / Architectures and efficiency / Attention and transformers
Text Compression-aided Transformer Encoding
Zuchao Li, Zhuosheng Zhang, Hai Zhao, Rui Wang, Kehai Chen, Masao Utiyama, Eiichiro Sumita
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 |
|---|---|---|---|
| add ETC Pipeline backbone fusion (BBF) | Transformer (base) · WMT14 EN-DE | BLEU: 27.24 -> 28.35 | Improved |
| add ITC Joint (BBF) instead of pipeline ETC (ablation) | Transformer (base) · WMT14 EN-DE | BLEU: 28.35 -> 28.57 | Improved |
| fuse backbone on decoder side (BDF) instead of encoder side (BEF) (ablation) | ETC Pipeline, Transformer (base) · WMT14 EN-DE | BLEU: 27.75 -> 28.14 | Improved |
| use meaningful compression (Semi-supervised) instead of random sampling (ablation) | ETC Pipeline BBF · WMT14 EN-DE | BLEU: 26.53 -> 28.35 | Improved |
| add AllText as extra source instead of compression (ablation) | ETC Pipeline BBF · WMT14 EN-DE | BLEU: 27.24 -> 27.24 | No clear effect |
| use F8W (first eight words) as compression (ablation) | ETC Pipeline BBF · WMT14 EN-DE | BLEU: 27.24 -> 27.40 | No clear effect |
| use unsupervised instead of supervised compression (ablation) | ETC Pipeline BBF · WMT14 EN-DE | BLEU: 27.80 -> 27.97 | Improved |
| use semi-supervised finetuning instead of unsupervised (ablation) | ETC Pipeline BBF · WMT14 EN-DE | BLEU: 27.97 -> 28.35 | Improved |
| use independent encoders instead of shared encoder (ablation) | ETC Pipeline BBF · WMT14 EN-DE | BLEU: 28.35 -> 28.50 | Mixed |
| add DAE pre-training instead of compression (ablation) | Transformer (base) · WMT14 EN-DE | BLEU: 27.24 -> 27.65 | Improved |
| train ETC jointly (ETC Joint) instead of pipeline (ablation) | ETC Pipeline BBF · WMT14 EN-DE | BLEU: 28.35 -> 28.04 | Hurt |
| use ETC trained on matching domain (TED) (ablation) | ETC Pipeline BBF · IWSLT15 TED test | BLEU: 30.76 -> 31.28 | Improved |
Closest papers by meaning
- ARC-Encoder: learning compressed text representations for large language models (NeurIPS 2026 · similarity 0.85)
- Language Modeling Is Compression (ICLR 2024 · similarity 0.85)
- Long Context Compression with Activation Beacon (ICLR 2025 · similarity 0.84)
- Distilling Sequential Computation in Transformer Language Models (NeurIPS 2026 · similarity 0.84)
- Compressing LLMs: The Truth is Rarely Pure and Never Simple (ICLR 2024 · similarity 0.84)
- Transformers Get Stable: An End-to-End Signal Propagation Theory for Language Models (ICML 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.
