Specialty Digest

Global English editionLearners’ Study Dashboard — in development
The Knowledge Library for All

Text Summarization

nounNatural Language Processingalso automatic summarization, document summarization, abstractive summarization

In one line

Text summarization is the task of automatically producing a shorter version of a text that keeps its most important information.

Text summarization, or automatic summarization, is the task of producing a shorter version of one or more texts that keeps their most important information. A good summary lets a reader grasp the main points quickly without reading the full document. The task is one of the long-standing problems in natural language processing, and it has become a mainstream feature of search engines, email tools, news apps and AI assistants.

Overview

Summarization has to balance three qualities. A summary should be faithful, meaning that it does not say anything the source does not support. It should be informative, covering the central content, and concise and readable. Achieving all three is difficult, and what counts as important depends on the reader and the purpose: a summary for an executive differs from one for a specialist.

Summarization requires natural language understanding to decide what matters, and often natural language generation to express it. Variants differ by input and output.

  • Single-document vs multi-document. One source, or many sources on the same topic.
  • Generic vs query-focused. A general overview, or a summary shaped around a question.
  • Indicative vs informative. A description of what a text covers, or a condensed version of its content.
  • Update summarization. Reporting only what is new compared with earlier material.

History

The first automatic approach is usually credited to Hans Peter Luhn of IBM, who in 1958 proposed scoring sentences by the frequency of significant words and selecting the highest ones as an abstract [1]. For decades, work followed this extractive path, adding features such as sentence position, title words and cue phrases, and later graph-based ranking such as TextRank. From the mid-2010s, neural sequence-to-sequence models made abstractive summarization practical, and pointer-generator networks added the ability to copy words from the source [3]. Transformer models then raised quality further, and today large language models summarize long documents through simple instructions.

Extractive summarization

Extractive methods build a summary by choosing existing sentences or phrases from the source and joining them. Typical steps are to split the text into sentences, score each sentence for importance, and select the top scorers while avoiding redundancy.

  • Frequency and position methods. Score by word statistics and where the sentence appears.
  • Graph-based ranking. Treat sentences as nodes linked by similarity and rank them, in the manner of PageRank.
  • Learned scoring. A supervised model predicts which sentences belong in a summary.

Extractive summaries are usually faithful because the sentences come directly from the source, but they can read awkwardly, since selected sentences may lack context or contain unresolved pronouns.

Abstractive summarization

Abstractive methods generate new text that may use words not present in the source, much as a person would paraphrase. Neural models read the document and generate the summary one token at a time. Results are more fluent and compact, and they can merge information from several sentences. The main risk is hallucination: the model may state details that are not in the source, or distort them. This makes faithfulness the central research problem for abstractive systems; see hallucination and grounding.

Many practical systems are hybrid, first selecting relevant passages and then rewriting them. Summarizing very long inputs, such as books or lengthy reports, requires strategies like splitting the text into pieces (chunking), summarizing each part and combining the results, because the input may exceed a model’s context window.

Evaluation

The most common automatic metric is ROUGE, a set of measures that count overlapping words and word sequences between a generated summary and human-written reference summaries [2]. It is cheap and standard, but it rewards surface similarity and can miss meaning, so a faithful paraphrase may score poorly while a wrong summary with overlapping words scores well. Newer metrics use embeddings or models to judge meaning, and some use language models as judges of faithfulness. Human evaluation of fluency, coverage and factual consistency is still considered the most reliable check.

Applications

  • News and media. Short digests of articles and alerts.
  • Search. Snippets and answer summaries in results, often combined with retrieval-augmented generation.
  • Meetings and email. Notes and action items, often starting from speech recognition transcripts.
  • Legal, financial and scientific work. Condensing contracts, filings and papers to help professionals triage.
  • Customer support. Summarizing long ticket histories and calls.
  • Healthcare. Drafting summaries of clinical records, which require clinician review.
  • Education and accessibility. Making long readings easier to approach.

Challenges and risks

  • Factual errors. Fluent summaries can contain mistakes or invented details.
  • Omission and bias. Important caveats may be dropped, and a summary can tilt the emphasis of a text.
  • Long documents. Quality can drop as input length grows.
  • Subjectivity. There is no single correct summary, and reference summaries differ among writers.
  • Domain language. Technical and legal text is harder to summarize accurately.
  • Over-reliance. Users who read only the summary may miss nuance, so important decisions should rest on the source.

Controlling summaries

Users often want summaries of a particular length, style or focus. Modern systems accept instructions such as a target number of sentences, a bullet-point format, a reading level or a question to focus on. For consistent results, developers use structured output templates that ask for fields like main points, decisions and open questions. Citations that point from each summary statement back to the passage it came from help readers verify claims, and are a practical way to reduce the harm of errors.

Summarizing different media

Summarization is not limited to plain text. Meeting and lecture summaries start from transcripts, which contain hesitations, interruptions and several speakers, so systems must first clean and segment the text. Summaries of scientific papers use the structure of sections such as abstract, methods and results. Multi-document summaries must detect repeated information and conflicts between sources. Video and image summarization extend the idea to other media, selecting key frames or writing descriptions, and they connect to image captioning and multimodal AI.

Choosing a method

When faithfulness is vital and wording is secondary, as in legal review, an extractive approach or an abstractive one with source citations is safer. When readability and brevity matter most, an abstractive model is more suitable, with checks on key facts and numbers. A small human-reviewed sample drawn from the output at regular intervals is a simple way to keep quality visible over time.

FAQs

What is text summarization? It is the automatic creation of a shorter text that keeps the key information of one or more longer documents.

What is the difference between extractive and abstractive summarization? Extractive methods select and join sentences already in the source. Abstractive methods write new sentences that paraphrase the content.

What is ROUGE? ROUGE is a family of metrics that measure how much a generated summary overlaps with human reference summaries, and it is widely used in summarization research.

Can AI summaries be wrong? Yes. Abstractive models in particular can introduce errors or invented details, so important summaries should be checked against the source.

What are the uses of text summarization? Common uses include news digests, meeting notes, search results, document review and customer support.

Key facts

  • Hans Peter Luhn proposed in 1958 an automatic method for creating literature abstracts by selecting sentences with frequent significant words1.
  • ROUGE, introduced by Chin-Yew Lin in 2004, is a family of metrics that score summaries by overlap with human reference summaries2.
  • The CNN/Daily Mail data set, adapted for summarization in 2015 and 2016, became a standard benchmark for abstractive models3.

Related concepts

Quick checkWhat is text summarization?Show answer

Text summarization is the automatic creation of a shorter text that keeps the main points of one or more longer documents, either by selecting key sentences (extractive) or by writing new ones (abstractive).

Sources

  1. H. P. Luhn, The Automatic Creation of Literature Abstracts, IBM Journal of Research and Development, https://doi.org/10.1147/rd.22.0159
  2. Chin-Yew Lin, ROUGE: A Package for Automatic Evaluation of Summaries, https://aclanthology.org/W04-1013/
  3. Abigail See, Peter Liu and Christopher Manning, Get To The Point: Summarization with Pointer-Generator Networks, https://arxiv.org/abs/1704.04368

Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue