presentsUNAIVERSE
The UN AI timeline

    Learn AI /Computer Vision and Language Technology /Summarisation

    Summarisation

    Also called: summarization, auto-summary.

    Detail

    In five seconds

    What it is

    Summarisation produces a shorter version of text or speech that keeps the main points and drops less important detail.

    In plain words

    It gives you the shorter version of a long text, keeping the main points and leaving out detail.

    An everyday example

    A team in Dakar, Lima and Bogotá shares ten pages of background notes before a call. One colleague reads them and says, "Here are the three points you need before we start: the budget changed, the deadline moved, and one partner has not replied yet." That briefing saves time because nobody has to hear every paragraph first. But if someone needs the exact wording of an agreement, they still have to open the full notes.

    Where the story stops working

    A human colleague may use wider context, shared history and judgement about what each listener needs. A summarisation system usually relies mainly on patterns in the source text and training data. It may miss nuance, uncertainty or implied meaning, and some systems copy sentences while others rewrite them, so the process is less flexible than the analogy suggests.

    Remember thisSummarisation compresses a source into its main points, either by selecting existing passages or generating new wording. The shorter version must still be checked for omissions and unsupported claims.

    Understand it

    Understand it

    Why it matters

    Summarisation matters because most people face more text than they can read in full. A good summary can help you triage documents, prepare for meetings, scan evidence, and find what deserves deeper review. For policy, communications and management work, that can save time and improve consistency. But the value depends on knowing its limits: a short version may omit conditions, caveats or minority views. If you understand the term, you are better able to ask when a summary is suitable, when the original must be checked, and what quality controls are needed.

    From long document to short summary

    Show how summarisation turns a long source into a shorter text by selecting and compressing the most important information.

    1. 01

      1. Source text enters

      The input may be a report, article, transcript or several related documents.

      Draw a tall document icon on the left with many horizontal lines. Add small tags such as report, article and transcript.

    2. 02

      2. Break into pieces

      The system analyses sentences or sections so it can compare them.

      Show the tall document splitting into stacked cards labelled sentence 1, sentence 2, sentence 3 and section notes.

    3. 03

      3. Find what matters most

      The system scores which parts seem central, repeated or informative.

      Place a ranking panel in the centre with bars of different lengths beside the cards. Highlight the top items in a darker colour.

    4. 04

      4. Produce a shorter version

      The output may copy the best sentences or rewrite them into a compact summary.

      Draw two small paths merging into one summary card on the right: one path labelled select sentences, one labelled rewrite. The final card has three short lines.

    A left-to-right flow diagram with four boxes: source text, split into units, rank important content, produce summary. Read from left to right to see what goes in, what the system does, and what comes out. Small callouts show that some systems copy key sentences while others rewrite them.

    Hover or tap each stage to reveal a plain-language note and an example of what is kept or dropped.

    Describe this without the diagram

    This diagram explains summarisation as a four-step flow. First, a long source text enters, such as a report, article or transcript. Second, the text is broken into smaller pieces like sentences or sections. Third, the system ranks those pieces to estimate which ones carry the main information. Fourth, it produces a shorter output. Some systems do this by selecting sentences from the original text. Others rewrite the content into new sentences. The key idea is that the summary starts from a source and tries to preserve its main points in less space.

    With reduced motion: A static left-to-right diagram shows four labelled boxes connected by arrows: source text, break into pieces, rank importance, produce summary. A note under the final box says the output may be extracted from the source or rewritten.

    Summarisation takes a long piece of content as input and returns a shorter version. The input might be an article, meeting notes, a report, a speech transcript, or several documents on the same topic. The output is not the whole original. It is a compressed version that tries to keep the main ideas.

    A system does this by finding what seems most important. In simple systems, that may mean selecting key sentences from the original text. In more advanced systems, it may mean generating new sentences that restate the main points. A transcript is a written record of spoken words. A model is a statistical system trained to detect patterns in data.

    Abstractive summarisation can use text generation, but the task is narrower: it starts from source material and aims to preserve its main points in less space. Translation changes language. Topic modelling groups themes rather than producing a readable short version.

    The main risk is loss or distortion. Important detail can be left out. A summary can overstate one point, miss uncertainty, or add wording that was not supported by the source. So a summary is useful for triage and first reading, but not a substitute for checking the original when accuracy matters.

    How “A colleague briefs you before a meeting” maps onto the real thing

    In the storyIn the AI system
    ten pages of background notesthe original long text or transcript
    the colleague's three-point briefingthe generated summary
    choosing only the most important updatesidentifying salient information
    checking the full notes for exact wordingverifying the source when precision matters

    A worked example

    Producing a concise summary of an implementing partner’s progress report for an oversight file.

    Starting from: Report excerpt: “Activities scheduled for the reporting period were completed except for the training package, which remains under review by the national counterpart. Field monitoring found inconsistent use of the revised reporting template. The partner issued additional guidance and plans refresher sessions. Supporting attendance records were submitted, while the expenditure annex remains pending internal reconciliation.”

    1. The system separates the excerpt into reported achievements, delays, corrective actions and outstanding evidence.
    2. It identifies the pending training package, inconsistent template use, planned refresher sessions and missing expenditure annex as salient points.
    3. An extractive system could select and reorder source sentences. An abstractive system instead generates shorter wording based on the same content.
    4. The auditor compares the generated summary with the excerpt to check that qualifications and pending items have not disappeared in compression.

    Result: Summary: “The partner reports that planned activities were largely completed, although the training package remains under review. Monitoring identified inconsistent use of the revised reporting template, prompting further guidance and planned refresher sessions. Attendance records were submitted, but the expenditure annex is still pending.”

    In a United Nations workplace

    An internal auditor reviewing an evaluation office’s oversight file receives a lengthy progress report from an implementing partner. Before close of business, the auditor uses an approved summarisation tool to prepare a short note for the file identifying reported achievements, delays, corrective actions and evidence still outstanding.

    The summary provides a first-pass map of the report and directs the auditor to sections requiring closer review. It does not replace the source document or establish whether the partner’s claims are supported.

    Before you try thisDo not upload unpublished reports, personal data or audit-sensitive material to a tool that has not been approved for them. Unauthorised disclosure could compromise confidentiality and due process. Check every finding, qualification and reference against the report before the summary enters the file.

    Where you may hear it

    • A vendor demonstration to an ICT governance board
    • A regional bureau front office clearing a briefing note
    • An evaluation office reviewing implementing-partner reporting
    • A desk officer preparing talking points before informals

    Avoid confusion

    Avoid confusion

    Often confused with

    Common misconceptions

    • A summary is just the first few sentences shortened.

      Useful summarisation usually tries to identify the most important information across the whole source. In many documents, the key point appears later, is repeated in several places, or depends on context, so simply trimming the start often fails.

    • If the summary reads well, it must be accurate.

      Fluent wording is not proof of faithfulness to the source. A system can produce a clear, polished summary that leaves out a condition, changes emphasis, or adds a detail not supported by the original text.

    • Summarisation always writes new sentences.

      Some systems extract existing sentences directly from the source, while others generate new wording. Both are called summarisation, but they have different strengths and risks.

    Quick check

    Which option best describes summarisation?

    Go deeper

    Go deeper

    The simple version

    It gives you the shorter version of a long text, keeping the main points and leaving out detail.

    The technical version

    Summarisation is a task in natural language processing that maps one or more source documents to a shorter text intended to preserve salient content. Two broad approaches are extractive summarisation, which selects and orders spans such as sentences from the source, and abstractive summarisation, which generates new wording conditioned on the source. Systems may use sentence scoring, attention mechanisms, encoder-decoder architectures, or large language models. Quality is often judged by faithfulness to the source, coverage of key information, coherence, and brevity, though automatic scores do not fully capture factual accuracy or usefulness.

    Reviewed external resources

    • EncyclopaediaAutomatic summarizationWikipedia

      This Wikipedia link gives a broad overview of automatic summarisation and suits readers who want a neutral starting point.

      Intermediatelink checked 2026-07-31

    • ArticleMachine Learning GlossaryGoogle

      This Google Machine Learning Glossary link helps readers who want short definitions of technical terms they may encounter around summarisation.

      Intermediatelink checked 2026-07-31

    • CourseMachine Learning Crash CourseGoogle

      This Google Machine Learning Crash Course link is useful for readers who want beginner-friendly background on machine learning concepts that support systems such as summarisation.

      Intermediatelink checked 2026-07-31

    • VideoVideo explanations: SummarisationYouTube search

      This YouTube search results page helps readers who prefer video explanations and want to compare several presentations of summarisation rather than rely on one source.

      search results, not one pagelink checked 2026-07-31

    • DiscussionCommunity discussion: SummarisationReddit search

      This Reddit search results page lets readers see community questions and informal discussion about summarisation, which may suit readers curious about common confusions and practical experiences.

      search results, not one pagelink checked 2026-07-31

    Where this came from

    Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

    The one thing to remember

    Summarisation compresses a source into its main points, either by selecting existing passages or generating new wording. The shorter version must still be checked for omissions and unsupported claims.