Research map / Large language models / Agents, code and math
Summarization Programs: Interpretable Abstractive Summarization with Neural Modular Trees
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 |
|---|---|---|---|
| Use best-of-five module outputs instead of one output per module (ablation) | SP-Search · CNN/DailyMail validation | ROUGE-2: 34.34 -> 40.11 | Helped |
| Add the paraphrase module (ablation) | SP-Search · CNN/DailyMail validation | ROUGE-2: 38.28 -> 40.11 | Helped |
| Add the compression module (ablation) | SP-Search · CNN/DailyMail validation | ROUGE-2: 37.44 -> 40.11 | Helped |
| Add the fusion module (ablation) | SP-Search · CNN/DailyMail validation | ROUGE-2: 31.30 -> 40.11 | Helped |
| Search for structured programs rather than randomly composing modules | Summarization Programs · CNN/DailyMail validation | ROUGE-2: 15.18 -> 40.11 | Helped |
| Train an SP generation model instead of randomly sampling programs | Joint SP model · CNN/DailyMail test | ROUGE-2: 6.48 -> 16.00 | Helped |
| Extract relevant source sentences before building the program | SP Model (Extract-and-Build) · CNN/DailyMail test | ROUGE-2: 16.40 vs 16.00 | Mixed |
| Use an SP generation model instead of PEGASUS | SP Model · CNN/DailyMail test | ROUGE-2: 21.55 vs 16.40 | Hurt |
| Show humans model-generated SPs before they simulate model reasoning | Human simulation · CNN/DailyMail human simulation study | ROUGE-2: 62.79 -> 66.80 | Helped |
| Increase the number of eligible source sentences from 4 to 5 (ablation) | SP-Search · CNN/DailyMail validation | ROUGE-2: 40.11 -> 41.99 | Helped |
| Increase maximum queue size from 20 to 30 (ablation) | SP-Search · CNN/DailyMail validation | ROUGE-2: 40.11 -> 40.57 | Helped |
| Increase maximum tree height from 2 to 3 (ablation) | SP-Search · CNN/DailyMail validation | ROUGE-2: 40.11 -> 43.81 | Mixed |
Closest papers by meaning
- Joint Parsing and Generation for Abstractive Summarization (AAAI 2020 · similarity 0.92)
- Controlling the Amount of Verbatim Copying in Abstractive Summarization (AAAI 2020 · similarity 0.89)
- Exploring Explainable Selection to Control Abstractive Summarization (AAAI 2021 · similarity 0.89)
- Distilling Structured Rationale from Large Language Models to Small Language Models for Abstractive Summarization (AAAI 2025 · similarity 0.87)
- Unsupervised Extractive Summarization with Learnable Length Control Strategies (AAAI 2024 · similarity 0.86)
- RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval (ICLR 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.