--- id: text-summarization title: "Text Summarization: From Extractive Methods to Abstractive LLM-Based Summarization" schema_type: article category: ai language: en confidence: high last_verified: "2026-05-24" created_date: "2026-05-24" generation_method: ai_assisted ai_models: - claude-4.5-sonnet derived_from_human_seed: true conflict_of_interest: none_declared is_live_document: false data_period: static completeness: 0.85 atomic_facts: - id: af-text-summarization-1 statement: >- ACM Computing Surveys (June 2025) published a systematic survey of text summarization tracing four eras: statistical methods (TF-IDF, TextRank, pre-2015) → neural sequence-to-sequence (Pointer-Generator, 2015-2019) → pre-trained language models (BART, PEGASUS, T5, 2019-2022) → LLM-based summarization (GPT-4, Claude, 2022-present) — finding that LLMs achieve ROUGE-L scores 10-15 points higher than fine-tuned BART on multi-document and conversational summarization, but struggle with factual consistency (hallucinating 3-8% of summary content for long documents). source_title: ACM Computing Surveys (2025) — A Systematic Survey of Text Summarization — doi:10.1145/3731445 source_url: https://dl.acm.org/doi/abs/10.1145/3731445 confidence: high - id: af-text-summarization-2 statement: >- arxiv (June 2024) and ScienceDirect (2024) surveys established long-document summarization as the key remaining challenge — for documents exceeding 10K tokens (legal briefs, scientific papers, meeting transcripts) — documenting that hierarchical models (discourse-aware, extract-then-abstract) and chunk-and-fuse approaches using LLMs with 100K+ context windows now achieve ROUGE-2 scores of 22-28 (vs. 18-22 for pre-LLM baselines) on PubMed and GovReport benchmarks. source_title: arxiv 2406.11289 (2024) / ScienceDirect NLP Journal (2024) — Text summarization surveys source_url: https://arxiv.org/abs/2406.11289 confidence: high primary_sources: - id: ps-text-summarization-1 title: "A Systematic Survey of Text Summarization: From Statistical Methods to Large Language Models" type: academic_paper year: 2025 institution: ACM Computing Surveys doi: 10.1145/3731445 url: https://dl.acm.org/doi/abs/10.1145/3731445 - id: ps-text-summarization-2 title: "A Systematic Survey of Text Summarization: From Statistical Methods to LLMs" type: academic_paper year: 2024 institution: arXiv url: https://arxiv.org/abs/2406.11289 known_gaps: - Factual consistency verification — detecting and correcting hallucinated summary content - Personalized summarization adapting to user background knowledge and information needs disputed_statements: [] secondary_sources: - title: "A Survey on Neural Text Summarization: Methods, Evaluation, and Applications" type: survey_paper year: 2024 authors: - multiple institution: ACM Computing Surveys url: https://doi.org/10.1145/3635100 - title: "Get To The Point: Summarization with Pointer-Generator Networks" type: conference_paper year: 2017 authors: - See, Abigail - Liu, Peter J. - Manning, Christopher D. institution: Stanford / ACL url: https://arxiv.org/abs/1704.04368 - title: "PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization" type: conference_paper year: 2020 authors: - Zhang, Jingqing - Zhao, Yao - Saleh, Mohammad - Liu, Peter J. institution: Google Research / ICML url: https://arxiv.org/abs/1912.08777 - title: "BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation" type: conference_paper year: 2019 authors: - Lewis, Mike - Liu, Yinhan - Goyal, Naman - et al. institution: Facebook AI Research / ACL url: https://arxiv.org/abs/1910.13461 updated: "2026-05-24" --- ## TL;DR Text summarization condenses documents into concise summaries while preserving key information. The field has evolved from simple sentence extraction to LLM-powered abstractive generation that rewrites content in its own words. The hard problems remain: summarizing book-length documents, ensuring factual accuracy, and adapting summaries to user needs. ## Core Explanation Two paradigms: (1) Extractive summarization — select and concatenate the most important sentences from the source text. Methods: TextRank (graph-based centrality), LexRank, BERT-based sentence scoring (BERTSUM, MatchSum). Pros: factually accurate (sentences are verbatim). Cons: disjointed, can't synthesize across sentences; (2) Abstractive summarization — generate new sentences that capture the essence. Architecture: encoder-decoder (BART, PEGASUS, T5) — encoder reads source, decoder generates summary. Training: teacher forcing on (document, reference summary) pairs from CNN/DailyMail, XSum, PubMed datasets. LLM-based: prompt-based generation ("Summarize the following article: [text]") with chain-of-thought for complex documents. ## Detailed Analysis Evaluation metrics: ROUGE (Recall-Oriented Understudy for Gisting Evaluation) — ROUGE-1 (unigram overlap), ROUGE-2 (bigram), ROUGE-L (longest common subsequence). ROUGE correlates moderately with human judgment but misses factual consistency. BERTScore (2020) uses contextual embeddings. Factuality metrics: SummaC, QAFactEval — verify summary claims against source document using NLI (Natural Language Inference). Long document summarization: the key bottleneck was context length — BART/PEGASUS limited to 1024 tokens. Solutions: (1) Hierarchical — encode sentences, then encode sentence representations (HIBERT); (2) Extractive then abstractive — select salient sentences first (30-50% compression), then generate abstractive summary; (3) LLM-based — models with 128K-1M token context windows (GPT-4, Claude, Gemini) directly process entire documents. Domain-specific summarization: medical (radiology reports → findings), legal (case law → holdings), scientific (full papers → structured abstracts). Meeting summarization: multi-speaker dialogue with topic segmentation. The ACM 2025 survey identifies the "summary faithfulness" problem as the most critical open challenge — 3-8% hallucination rate is unacceptable for high-stakes domains. ## Further Reading - BART: Denoising Sequence-to-Sequence Pre-training (Lewis et al., ACL 2020) - PEGASUS: Pre-training with Extracted Gap-sentences (Zhang et al., ICML 2020) - Hugging Face Summarization Pipeline