Research map / Architectures and efficiency / Attention and transformers
DCN+: Mixed Objective And Deep Residual Coattention for Question Answering
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 deep residual coattention (ablation) | DCN+ · SQuAD dev | EM: 73.1% -> 74.5% (-1.4% when removed) | Helped |
| Add deep residual coattention (ablation) | DCN+ · SQuAD dev | F1: 81.5% -> 83.1% (-1.6% when removed) | Helped |
| Add mixed cross-entropy and self-critical policy objective (ablation) | DCN+ · SQuAD dev | EM: 73.8% -> 74.5% (-0.7% when removed) | Helped |
| Add mixed cross-entropy and self-critical policy objective (ablation) | DCN+ · SQuAD dev | F1: 82.1% -> 83.1% (-1.0% when removed) | Helped |
| Add sparse mixture of experts decoder layer (ablation) | DCN+ · SQuAD dev | EM: 74.0% -> 74.5% (-0.5% when removed) | Helped |
| Add sparse mixture of experts decoder layer (ablation) | DCN+ · SQuAD dev | F1: 82.4% -> 83.1% (-0.7% when removed) | Helped |
| Combine cross-entropy supervision with self-critical F1 policy learning instead of using cross-entropy alone | DCN+ · SQuAD training | F1 | Helped |
| Use self-critical policy learning without the cross-entropy task | DCN+ · SQuAD training | policy convergence | Hurt |
| Use DCN+ rather than DCN with CoVe | Question answering models · SQuAD dev | EM: +3.2% | Helped |
| Use DCN+ rather than DCN with CoVe | Question answering models · SQuAD dev | F1: +3.2% | Helped |
| Use deep residual coattention and mixed objective | DCN+ · SQuAD long questions | performance | Helped |
Closest papers by meaning
- Dynamic Coattention Networks For Question Answering (ICLR 2017 · similarity 0.91)
- Coarse-grain Fine-grain Coattention Network for Multi-evidence Question Answering (ICLR 2019 · similarity 0.88)
- Beyond RNNs: Positional Self-Attention with Co-Attention for Video Question Answering (AAAI 2019 · similarity 0.85)
- Generative Question Answering: Learning to Answer the Whole Question (ICLR 2019 · similarity 0.85)
- FRAMES-VQA: Benchmarking Fine-Tuning Robustness across Multi-Modal Shifts in Visual Question Answering (CVPR 2025 · similarity 0.83)
- GLEN: Generalized Focal Loss Ensemble of Low-Rank Networks for Calibrated Visual Question Answering (AAAI 2025 · similarity 0.83)
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.