Research map / Vision-language and multimodal / Visual question answering
S-Net: From Answer Extraction to Answer Synthesis for Machine Reading Comprehension
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 answer synthesis model on top of extracted evidence (ablation) | S-Net · MS-MARCO dev | ROUGE-L: 42.23 -> 45.95 | Helped |
| add answer synthesis model on top of extracted evidence (ablation) | S-Net · MS-MARCO test | BLEU-1: 44.08 -> 43.78 | Hurt |
| add passage ranking multi-task objective (ablation) | S-Net extraction · MS-MARCO dev | ROUGE-L (extraction): 39.62 -> 42.23 | Helped |
| add passage ranking multi-task objective (ablation) | S-Net extraction · MS-MARCO dev | passage selection P@1: 34.6 -> 38.9 | Helped |
| joint learning of passage ranking and span prediction instead of two separate steps (ablation) | S-Net extraction · MS-MARCO dev | passage selection P@1: 28.3 -> 38.9 | Helped |
| apply synthesis to other extraction models (BiDAF, Prediction) | BiDAF, Prediction · MS-MARCO dev | ROUGE-L: 34.89 -> 38.73; 37.54 -> 41.55 | Helped |
| ensemble 13 extraction models | S-Net (Extraction) · MS-MARCO test | ROUGE-L: 41.45 -> 42.92 | Helped |
| add synthesis for questions whose answers exactly match passage spans (ablation) | S-Net · MS-MARCO dev (upper bound = 1.0) | ROUGE-L: 50.74 -> 49.59 | Hurt |
| add synthesis for questions with upper bound <0.2 (e.g., yes/no) (ablation) | S-Net · MS-MARCO dev | ROUGE-L: 0.01 -> 49.18 | Helped |
| two-step extraction-then-synthesis instead of end-to-end generation with selected passage (ablation) | answer generation · MS-MARCO dev | ROUGE-L: 37.70 -> 45.95 | Helped |
| end-to-end matching + seq2seq instead of two-step (ablation) | answer generation · MS-MARCO dev | ROUGE-L: 6.28 | Hurt |
Closest papers by meaning
- Machine Comprehension Using Match-LSTM and Answer Pointer (ICLR 2017 · similarity 0.89)
- Multi-Matching Network for Multiple Choice Reading Comprehension (AAAI 2019 · similarity 0.88)
- Multi-Task Learning with Generative Adversarial Training for Multi-Passage Machine Reading Comprehension (AAAI 2020 · similarity 0.88)
- Read + Verify: Machine Reading Comprehension with Unanswerable Questions (AAAI 2019 · similarity 0.87)
- GAAMA 2.0: An Integrated System That Answers Boolean and Extractive Questions (AAAI 2023 · similarity 0.87)
- Generating Well-Formed Answers by Machine Reading with Stochastic Selector Networks (AAAI 2020 · similarity 0.87)
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.