Automatic summarization¶

Automatic production of a shorter text retaining salient information from source material.

Version: v1 · 2026-09-28 · History
Domain-specific #: 8084
Domain group: Applied Sciences & Engineering
Origin domain: Computer Science & Software Engineering
Subdomain: Natural Language Processing → Computer Science & Software Engineering

Core Idea¶

Automatic summarization is the computational production of a shorter representation that preserves information judged important for a task, audience, or query. The input may be one document, many documents, an image collection, audio, video, or mixed data; the output may be text, key phrases, representative images, key frames, or selected segments. A summarizer must therefore define both a selection objective—importance, coverage, novelty, relevance, chronology, or user need—and a compression constraint. Shortening alone is insufficient: deleting arbitrary content produces a smaller object but not a defensible summary.

How would you explain it like I'm…

The Short-Version Maker

When you tell a friend about a long movie in just a few sentences, you keep the most important parts and skip the rest. Automatic summarization is a computer doing that with long writing, pictures, or videos. Just cutting off random pieces isn't a summary; the short version has to keep what matters.

Computer-Made Summaries

Automatic summarization is when a computer takes something big — an article, many articles, a video, or a pile of pictures — and makes a much shorter version that keeps the important information. One way is to pick out the best existing pieces, like key sentences or video clips. Another way is to write something new in its own words, which can be shorter and smoother but might accidentally say things the original didn't. A summary can aim to cover everything in general or focus on a particular question. A good summary isn't just short and smooth-sounding; it has to be true to the original and useful for the reader.

Task-Guided Automatic Summaries

Automatic summarization is the computer production of a shorter representation that keeps the information judged important for a task, audience, or question. Inputs can be one document, many documents, images, audio, or video; outputs can be text, key phrases, representative images, key frames, or clips. A summarizer needs a goal for what to keep (importance, coverage, novelty, relevance, chronology, user needs) and a length limit — just deleting content isn't summarizing. Extractive methods select existing pieces, which is traceable and less likely to invent things but can be repetitive or choppy. Abstractive methods write new text, which can combine ideas compactly but risks unsupported or wrong claims. Evaluation depends on the source, task, and readers: fluent text isn't necessarily faithful, and word-overlap scores can punish a correct paraphrase while rewarding copied but unimportant text.

Automatic summarization is the computational transformation of a larger information object into a smaller one that preserves information judged important for a task, audience, or query. Inputs range over single or multiple documents, image collections, audio, video, or mixed data; outputs may be text, key phrases, representative images, key frames, or selected segments. A summarizer must specify a selection objective (importance, coverage, novelty, relevance, chronology, or user need) and a compression constraint; arbitrary deletion yields a smaller object but not a defensible summary. Extractive methods select existing units (sentences, phrases, frames, shots), gaining traceability and lower fabrication risk at the cost of redundancy or incoherent transitions; abstractive methods generate new representations by paraphrasing, combining, and reorganizing, enabling syntheses no single source unit states but requiring management of factual consistency, attribution, and unsupported inference. Hybrid and human-aided workflows combine machine proposal with human revision, and generic summarization differs from query-focused summarization, which conditions importance on a stated need. Evaluation is relational to source, task, length budget, and users: fluency is not faithfulness, and lexical-overlap metrics can penalize correct abstraction while rewarding copied but unimportant text. Multi-document settings add contradiction, temporal updating, source diversity, and redundancy, and video synopsis that synthesizes composite frames differs from key-shot selection.

Scope of Application¶

Single-document text. Extractive, abstractive, or hybrid systems reduce articles, reports, transcripts, and records for defined users.
Multi-document synthesis. Redundant and conflicting sources require provenance, temporal ordering, and explicit handling of disagreement.
Query-focused output. Selection is conditioned on a user's information need rather than general salience alone.
Conversation and meeting summaries. Decisions, participants, uncertainty, and action items must remain attributable to the source interaction.
Audio, image, and video. Temporal or visual segments can be selected or described under modality-specific fidelity criteria.

Clarity¶

Automatic summarization makes shortening answerable to an information objective and a compression constraint. It distinguishes extractive selection from abstractive generation, single-document from multi-document synthesis, and generic coverage from query- or audience-focused relevance. The term prevents a shorter output from counting as a summary merely because content was deleted.

Manages Complexity¶

Automatic summarization turns a large source collection into an optimization over coverage, relevance, novelty, redundancy, coherence, faithfulness, and length. The analyst specifies the audience or query, information units, compression budget, and error costs rather than judging ‘shorter’ as a single property. Extractive and abstractive branches trade traceability against expressive compression; single- and multi-document settings add different redundancy and contradiction problems.

Abstract Reasoning¶

Selection move. From source units and an importance or relevance objective, infer which content deserves limited summary capacity. Compression move. Combine or rephrase units only when their source support and distinctions survive; otherwise prefer traceable extraction. Coverage move. Compare the output against required topics and redundancy to infer what information was lost or overrepresented. Faithfulness move. Trace every asserted fact to the input and reject fluent additions unsupported there. Boundary move.

Knowledge Transfer¶

Within the home domain. Automatic summarization transfers across news, scientific literature, meetings, legal documents, dialogue, and multimedia when a system selects or generates a shorter representation preserving task-relevant content. Source grounding, compression, salience, redundancy, coherence, and evaluation retain operational roles. Beyond the home domain (C — computational instrument). It applies literally wherever an input representation and summary objective are defined. Its boundary is over-reading: brevity does not guarantee factuality, coverage, neutrality, or suitability for a user; reference metrics do not exhaust quality. Human memory, abstraction, and institutional reporting may summarize, but are not automatic summarization unless an algorithm performs the transformation.

Relationships to Other Abstractions¶

Current abstraction Automatic summarization Domain-specific

Parents (1) — more general patterns this builds on

Automatic summarization is a kind of Representation Prime

Automatic summarization is a domain-specific kind of Representation: Automatic production of a shorter text retaining salient information from source material.

Hierarchy path (1) — routes to 1 parentless root

Automatic summarization → Representation → Abstraction

Neighborhood in Abstraction Space¶

Automatic summarization sits in a moderately populated region (57^th percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.

Family — Unclustered & Miscellaneous (2551 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-08