New framework improves factual consistency in abstractive summarization

Researchers have developed a new framework for abstractive summarization that decouples the generation and selection processes to improve factual consistency and control summary length. This modular approach uses a pretrained generator to create multiple candidate summaries, which are then evaluated and selected by a combinatorial process. Experiments on various datasets, including CNN/DailyMail and Multi-News, demonstrated consistent improvements in factuality and source-grounding metrics, with human evaluations indicating higher perceived quality. AI

IMPACT This research offers a novel approach to enhance the reliability and faithfulness of AI-generated summaries, potentially improving applications that rely on text summarization.

RANK_REASON The item is a research paper detailing a new method for abstractive summarization. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX
• CNN/DailyMail
• DagsHub
• FaithBench
• Gotit.pub
• Hugging Face
• Multi-News
• ScienceCast
• TofuEval

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →