Research map / Label-efficient and robust learning / Recognition and long-tail
MeanSum: A Neural Model for Unsupervised Multi-Document Abstractive Summarization
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 |
|---|---|---|---|
| Train MeanSum on reviews using its unsupervised summarization objective rather than leaving the pretrained model untrained | MeanSum · Yelp reference summaries test split | ROUGE-1: 21.2206140518 -> 28.8621865823 | Helped |
| Use unsupervised abstractive summarization instead of extractive summarization | Yelp · Yelp | sentiment accuracy: 42.94511675834656 -> 51.74971847115324 | Helped |
| Use unsupervised abstractive summarization instead of the extractive baseline | MeanSum · Yelp reference summaries test split | ROUGE-1: 24.6140384677 -> 28.8621865823 | Helped |
| Use unsupervised abstractive summarization instead of extractive summarization | Yelp · Yelp | word overlap: 28.593489521767257 -> 26.09 | Hurt |
| Remove pre-trained language-model initialization (ablation) | MeanSum · Yelp reference summaries test split | ROUGE-1: 28.8621865823 -> 26.1582704283 | Hurt |
| Remove pre-trained language-model initialization (ablation) | MeanSum · Yelp | sentiment accuracy: 51.74971847115324 -> 48.97116806308535 | Hurt |
| Remove the auto-encoder (ablation) | MeanSum · Yelp | natural-language output | Hurt |
| Replace the auto-encoder reconstruction objective with reconstruction-cycle loss (ablation) | MeanSum · Yelp reference summaries test split | ROUGE-1: 28.8621865823 -> 25.229314882 | Hurt |
| Replace summary-encoding cosine loss with early cosine loss on the mean review representation (ablation) | MeanSum · Yelp reference summaries test split | ROUGE-1: 28.8621865823 -> 14.350689634 | Hurt |
| Replace summary-encoding cosine loss with early cosine loss on the mean review representation (ablation) | MeanSum · Yelp | NLL: 1.1897457717087458 -> 1.7109228513580406 | Hurt |
| Do not share weights between the review and summary encoders (ablation) | MeanSum · Yelp reference summaries test split | ROUGE-1: 28.8621865823 -> 29.346102175 | No clear effect |
| Do not share weights between the review and summary decoders (ablation) | MeanSum · Yelp | natural-language output | Hurt |
Closest papers by meaning
- Summarization Programs: Interpretable Abstractive Summarization with Neural Modular Trees (ICLR 2023 · similarity 0.89)
- Abstractive Text Summarization by Incorporating Reader Comments (AAAI 2019 · similarity 0.89)
- Towards Improving Faithfulness in Abstractive Summarization (NeurIPS 2022 · similarity 0.89)
- Unsupervised Opinion Summarization with Content Planning (AAAI 2021 · similarity 0.89)
- Learning to Summarize by Learning to Quiz: Adversarial Agentic Collaboration for Long Document Summarization (ICLR 2026 · similarity 0.86)
- Distilling Structured Rationale from Large Language Models to Small Language Models for Abstractive Summarization (AAAI 2025 · 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.