Research map / Large language models / Language, safety and interpretability
MacNet: Transferring Knowledge from Machine Comprehension to Sequence-to-Sequence Models
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 |
|---|---|---|---|
| Add MacNet's transferred MC layers and focal loss | Baseline NMT · WMT14 and WMT15 En↔De | BLEU: 22.1 -> 24.2; 26.0 -> 28.1; 24.5 -> 26.3; 27.5 -> 29.4 | Helped |
| Add both transferred MC encoding and modeling layers, without focal loss (ablation) | Baseline NMT · WMT14 and WMT15 En↔De | BLEU: 22.1 -> 23.4; 26.0 -> 27.3; 24.5 -> 25.6; 27.5 -> 28.5 | Helped |
| Add the transferred MC encoding layer only (ablation) | Baseline NMT · WMT14 and WMT15 En↔De | BLEU: 22.1 -> 23.2; 26.0 -> 27.0; 24.5 -> 25.3; 27.5 -> 28.3 | Helped |
| Add the transferred MC modeling layer only (ablation) | Baseline NMT · WMT14 and WMT15 En↔De | BLEU: 22.1 -> 22.4; 26.0 -> 26.4; 24.5 -> 24.8; 27.5 -> 27.8 | Helped |
| Replace pretrained MC knowledge with random initialization (ablation) | Baseline NMT with transferred layers · WMT14 and WMT15 En↔De | BLEU: 23.4 -> 21.6; 27.3 -> 25.6; 25.6 -> 24.2; 28.5 -> 27.0 | Hurt |
| Add MacNet's transferred MC layers | Pointer-Generator · CNN/Daily Mail test | ROUGE-1/2/L: 39.69 -> 40.87; 17.26 -> 18.02; 36.38 -> 37.54 | Helped |
| Add MacNet's transferred MC layers | Pointer-Generator · Gigaword test | ROUGE-1/2/L: 36.44 -> 37.97; 17.26 -> 18.16; 33.92 -> 34.93 | Helped |
| Add the transferred MC encoding layer only (ablation) | Pointer-Generator · CNN/Daily Mail and Gigaword | ROUGE-1/2/L: 39.69 -> 40.38; 17.26 -> 17.75; 36.38 -> 37.24; 36.44 -> 37.30; 17.26 -> 17.83; 33.92 -> 34.41 | Helped |
| Add the transferred MC modeling layer only (ablation) | Pointer-Generator · CNN/Daily Mail and Gigaword | ROUGE-1/2/L: 39.69 -> 39.92; 17.26 -> 17.58; 36.38 -> 36.65; 36.44 -> 36.85; 17.26 -> 17.45; 33.92 -> 34.12 | Helped |
| Use context-to-query attention instead of BiDAF (ablation) | MacNet NMT · WMT14 De→En | BLEU: 27.5 -> 25.1 | Hurt |
| Use query-to-context attention instead of BiDAF (ablation) | MacNet NMT · WMT14 De→En | BLEU: 27.5 -> 25.3 | Hurt |
| Add self-attention to BiDAF (ablation) | MacNet NMT · WMT14 De→En | BLEU: 27.5 -> 27.4 | Hurt |
Closest papers by meaning
- SG-Net: Syntax-Guided Machine Reading Comprehension (AAAI 2020 · similarity 0.87)
- Machine Comprehension Using Match-LSTM and Answer Pointer (ICLR 2017 · similarity 0.87)
- A Graph Fusion Approach for Cross-Lingual Machine Reading Comprehension (AAAI 2023 · similarity 0.86)
- Structured Neural Summarization (ICLR 2019 · similarity 0.86)
- Deliberation Networks: Sequence Generation Beyond One-Pass Decoding (NeurIPS 2017 · similarity 0.86)
- SeqGPT: An Out-of-the-Box Large Language Model for Open Domain Sequence Understanding (AAAI 2024 · 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.