Research map / Vision-language and multimodal / Human and embodied modeling
Improving Faithfulness in Abstractive Text Summarization with EDUs Using BART (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes guiding BART-based abstractive summarization (via GSum) with important elementary discourse units (EDUs) extracted by DiscoBERT to improve faithfulness on CNN/DailyMail.
Problem
Abstractive summarization often produces unfaithful (inaccurate or misleading) summaries relative to the source document.
Approach
Uses important EDUs identified by a discourse-based extractor (DiscoBERT) as guidance signals fed into the GSum framework instantiated with BART, alone or combined with highlighted sentences, to steer generation toward faithful content.
Key results
- BART+EDUs(G) reaches ROUGE-1 0.54, ROUGE-2 0.31, ROUGE-L 0.50, versus 0.44/0.21/0.41 for plain BART.
- BART+EDUs(G) scores QAGS 0.084 vs 0.049 for BART, while BART+Sents,EDUs(G) gets the best entailment (0.929) and QAGS (0.093).
- BART+EDUs(G) has lower SentSim (0.738) than BART+Sents(G) (0.85), an inconsistency with the claimed gains.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Guide BART with important EDUs instead of highlighted sentences | BART summarization · CNN/Daily Mail | ROUGE-1: 0.45 -> 0.54 | Helped |
| Guide BART with important EDUs instead of highlighted sentences | BART summarization · CNN/Daily Mail | ROUGE-2: 0.22 -> 0.31 | Helped |
| Guide BART with important EDUs instead of highlighted sentences | BART summarization · CNN/Daily Mail | ROUGE-L: 0.42 -> 0.50 | Helped |
| Add important EDU guidance to BART | BART summarization · CNN/Daily Mail | ROUGE-1: 0.44 -> 0.54 | Helped |
| Add important EDU guidance to BART | BART summarization · CNN/Daily Mail | ROUGE-2: 0.21 -> 0.31 | Helped |
| Add important EDU guidance to BART | BART summarization · CNN/Daily Mail | ROUGE-L: 0.41 -> 0.50 | Helped |
| Add both sentence and EDU guidance rather than EDU guidance alone (ablation) | BART summarization · CNN/Daily Mail | Entail: 0.916 -> 0.929 | Helped |
| Add both sentence and EDU guidance rather than EDU guidance alone (ablation) | BART summarization · CNN/Daily Mail | QAGS: 0.084 -> 0.093 | Helped |
| Add both sentence and EDU guidance rather than EDU guidance alone (ablation) | BART summarization · CNN/Daily Mail | SentSim: 0.738 -> 0.854 | Helped |
Surprising findings
- Using EDUs alone yields lower SentSim (0.738) than sentence guidance (0.85) and lower entailment (0.916 vs 0.925), despite higher ROUGE.
Limitations
- Evaluation limited to CNN/DailyMail and a short student abstract.
- Test-time guidance relies on predicted (non-oracle) EDUs, while training uses oracle guidance.
- Reported confidence intervals and experimental details are sparse.
Experimental setup
Datasets: CNN/Daily Mail
Benchmarks: CNN/Daily Mail
Metrics: ROUGE-1 ROUGE-2 ROUGE-L BERTScore textual entailment QAGS SentSim
Baselines: BART GSum (highlighted sentences) CLIFF SEASON
Builds on: GSum BART DiscoBERT MatchSum
Closest papers by meaning
- Towards Improving Faithfulness in Abstractive Summarization (NeurIPS 2022 · similarity 0.86)
- Faithful to the Original: Fact Aware Neural Abstractive Summarization (AAAI 2018 · similarity 0.85)
- PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization (ICML 2020 · similarity 0.85)
- Exploring Human-Like Reading Strategy for Abstractive Text Summarization (AAAI 2019 · similarity 0.84)
- Automatic Summarization of Long Documents (Student Abstract) (AAAI 2025 · similarity 0.84)
- Document Summarization with Conformal Importance Guarantees (NeurIPS 2025 · similarity 0.82)
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.