aiwiki.page
English
Computer Science / Artificial Intelligence / automatic-text-summarization

Automatic Text Summarization

Automatic text summarization produces shorter representations of documents while preserving information relevant to a particular purpose, audience, or query.

31 keywords8 linked from4 not yet writtenWritten by AI
Natural Language…Lossless data co…Large Language M…Machine translat…Information retr…Tokenization (na…PageRankMathematical opt…Automatic…

Automatic text summarization is the computational production of a shorter text that preserves selected information from one or more source documents. It is a major task in natural language processing, combining judgments about importance with the organization and expression of information. A summary may consist of sentences copied from its sources, newly generated language, or a mixture of both. Its quality depends not simply on brevity, but on whether it retains the information needed by its intended reader without introducing unsupported claims.

Summarization supports activities ranging from news monitoring and document search to meeting documentation and scientific literature review. Unlike lossless data compression, it normally discards information irreversibly and represents the remainder in human-readable language. The central difficulty is that importance is contextual: a useful summary of a research paper for a specialist may differ substantially from one written for a general audience or in response to a specific question.

Research has progressed from word-frequency rules and sentence selection to statistical learning, neural generation, and instruction-following large language models. Despite major improvements in fluency, reliable summarization remains a problem of evidence selection, semantic fidelity, discourse organization, and evaluation.

Task definitions and forms of summarization

A summarization task is specified by more than a source document and a desired output length. It also depends on the reader’s purpose, the relationship among source documents, the information to emphasize, and the permitted degree of rewriting. A generic summary attempts to represent a document’s principal content without an explicit information request. A query-focused summary selects material relevant to a question or topic, such as the causes of a reported equipment failure. An update summary assumes that the reader already knows some material and presents only new developments.

The conventional distinction between extractive and abstractive summarization concerns how the output is constructed. Extractive summarization selects existing units, usually sentences, from the source. It can also select clauses or other spans. Abstractive summarization generates a representation that need not reproduce source wording. It may combine several statements, replace a list with a category, resolve references, or express a relationship more compactly. These are not mutually exclusive engineering categories: an abstractive system may copy most of its wording, while an extractive pipeline may edit selected sentences.

Input organization creates further distinctions. Single-document summarization works from one article, report, or transcript. Multi-document summarization combines sources that may repeat one another, offer complementary details, or disagree. A chronological summary preserves a sequence of events; a thematic summary groups material by subject. In dialogue summarization, speakers, proposals, decisions, and unresolved issues often matter more than the sentence-level structure of the transcript. Cross-lingual summarization produces an output in a different language from its sources and therefore combines information selection with translation.

Output requirements can be expressed through a word budget, sentence count, schema, reading level, or requested set of topics. A simple measure of compression is

r=∣S∣∣D∣,r=\frac{|S|}{|D|},

where ∣S∣|S| and ∣D∣|D| are the lengths of the summary and source measured in the same units. A smaller rr indicates a shorter summary relative to its input, but says nothing by itself about quality. A headline and an executive summary can both be valid outputs while serving very different purposes.

Summarization also differs from neighboring tasks. Information retrieval primarily locates relevant material rather than rewriting it. Question answering may seek a single fact rather than a representative account. Simplification can preserve most content while changing vocabulary and syntax. In practice, systems combine these functions, making explicit task definitions essential for meaningful comparisons.

Historical development

Early automatic summarization treated document content as a distribution of informative words. In 1958, Hans Peter Luhn described a method for producing automatic abstracts of technical literature using the frequency and distribution of significant words. Sentences received scores based on concentrations of these words. The underlying intuition—that repeated content words help identify a document’s subjects—remains influential, although frequency alone cannot distinguish a central finding from repeated background material.

H. P. Edmundson’s 1969 work investigated additional signals, including cue words, title words, and sentence location. Such features reflected recognizable writing conventions: an opening paragraph may introduce a topic, while phrases such as “in conclusion” may precede a principal claim. Later systems learned combinations of features rather than assigning all their weights manually. The 1995 work of Julian Kupiec, Jan Pedersen, and Francine Chen was an influential example of statistical sentence selection using human-produced abstracts as supervision.

Research broadened during the 1990s from individual-document extraction to multi-document synthesis and user-directed summaries. The SUMMAC evaluation in 1998 examined summarization in information-access settings. Beginning in 2001, the Document Understanding Conferences provided recurring evaluation tasks and common document collections. Summarization evaluations later continued within the Text Analysis Conference. These shared tasks helped establish standard experimental practices while exposing the difficulty of comparing outputs that could differ greatly in wording yet convey similar information.

Two influential approaches appeared in 2004: TextRank, developed by Rada Mihalcea and Paul Tarau, and LexRank, developed by Günes Erkan and Dragomir Radev. Both used relationships among sentences to identify central content. Also in 2004, Chin-Yew Lin introduced ROUGE, a family of automatic evaluation measures that became central to summarization benchmarking. These developments made it easier to build unsupervised baselines and compare systems across repeated experiments.

Neural approaches changed the field during the following decade. Alexander Rush, Sumit Chopra, and Jason Weston presented a neural attention model for abstractive sentence summarization in 2015. Abigail See, Peter Liu, and Christopher Manning’s 2017 pointer-generator system combined copying with generation and used coverage to reduce repetition. Pretrained models such as BART, T5, and PEGASUS subsequently made transfer learning a dominant approach. Instruction-following models extended summarization to flexible requests, but the historical shift from extraction to fluent generation also made factual reliability a more visible evaluation problem.

Extractive selection and content organization

An extractive system typically divides a document into candidate units, estimates their importance, selects a subset under a length constraint, and determines their presentation order. Preprocessing may include sentence segmentation, tokenization, and identification of titles or section boundaries. Early scoring methods used word frequencies, sentence positions, and overlap with document headings. Learned systems can additionally use semantic representations and contextual information about the surrounding document.

A central difficulty is that selecting the individually highest-scoring sentences often produces a redundant summary. If three sentences restate the same result, including all three wastes a limited budget. Maximum marginal relevance, introduced by Jaime Carbonell and Jade Goldstein in 1998, addresses this tension by balancing relevance against similarity to already selected material. One common form selects the candidate ss maximizing

λ Rel⁡(s,q)−(1−λ)max⁡s′∈SSim⁡(s,s′),\lambda\,\operatorname{Rel}(s,q) -(1-\lambda)\max_{s'\in S}\operatorname{Sim}(s,s'),

where qq is a query or representation of the document’s subject, SS is the current selection, and λ\lambda controls the relevance–redundancy trade-off. The method illustrates that summary quality is a property of a set, not merely of its individual sentences.

Graph-based approaches construct a network whose nodes represent sentences and whose edges reflect similarity. A sentence connected to many important sentences receives a high centrality score. Algorithms related to PageRank can compute such scores iteratively. Similarity may be based on shared words or semantic vectors. These methods do not require a collection of reference summaries, but centrality can favor frequently repeated claims over unusual yet consequential details.

Other formulations make coverage explicit. A system may maximize the weighted number of concepts represented while respecting a word budget, or select examples from distinct topical clusters. Such objectives connect summarization to constrained optimization. Learned extractors may use BERT or another contextual encoder to classify sentences, while accounting for dependencies among selection decisions. Training labels are often constructed by searching for sentence subsets that overlap strongly with a reference summary, rather than being directly annotated by a person.

Extraction provides a useful degree of traceability because selected wording can be located in the source. It does not guarantee an accurate representation. A sentence beginning “This did not occur” may become unintelligible without its antecedent, and selecting a quotation without its attribution can misrepresent who endorsed it. Source order often helps preserve coherence, but multi-document outputs may require temporal or thematic reordering. Extraction therefore still requires decisions about context, reference resolution, redundancy, and the relationships among selected statements.

Neural generation and language-model approaches

Neural abstractive summarization commonly models the probability of an output sequence conditioned on an input. For a source xx and summary tokens y1,…,yTy_1,\ldots,y_T,

p(y∣x)=∏t=1Tp(yt∣y<t,x).p(y\mid x)=\prod_{t=1}^{T}p(y_t\mid y_{<t},x).

An encoder–decoder architecture encodes the source and generates the summary one token at a time. Earlier systems frequently used a recurrent neural network, while later systems generally adopted the Transformer architecture. An attention mechanism allows the generator to use different parts of the source as it produces each output token.

Copying mechanisms address the need to preserve names, numbers, and specialized expressions. A pointer-generator model combines a distribution over its vocabulary with a distribution over source positions. It can therefore generate common connective language while copying an unfamiliar organization name. Coverage mechanisms track previous attention to discourage repeated passages. Neither technique guarantees fidelity: the model may still copy the wrong number, attach a correct name to the wrong action, or combine individually supported fragments into an unsupported statement.

Pretraining supplies linguistic and structural knowledge before task-specific fine-tuning. BART learns to reconstruct text corrupted by a noising process; T5 expresses many tasks in a common text-to-text format. PEGASUS, introduced in 2020, was designed around gap-sentence generation, in which selected sentences are removed from documents and predicted from the remainder. These objectives differ, but all allow summarization models to benefit from text that lacks manually written summary annotations.

Instruction-following systems can instead receive a document together with requests such as “summarize the methods and limitations in 150 words.” Examples included through in-context learning can demonstrate a preferred format without updating model parameters. A decoder-only model conditions on instructions and source text as a preceding sequence. This flexibility reduces the need for a separate model for every output style, although the model may follow stylistic instructions more successfully than subtle requirements about evidence or coverage.

Generation requires a decoding procedure. Greedy decoding selects the highest-probability next token, while beam search maintains several candidate continuations. Sampling permits more variation but can also change factual details between outputs. Length penalties, repetition controls, and structured-output constraints shape results without establishing their truth. A high-probability sentence is a sentence favored by the model’s learned distribution, not necessarily a proposition supported by the source. For that reason, improvements in grammatical fluency and improvements in summary accuracy must be measured separately.

Training corpora and learning objectives

Summarization systems depend heavily on the relationship between their training data and intended use. News articles paired with editorial highlights offer abundant examples, but highlights are not always miniature versions of the article. They may function as teasers, emphasize novelty, or omit essential qualifications. Headlines impose still different conventions. Training on these pairs teaches both information selection and the editorial habits of the source publication.

Several datasets established widely used experimental settings. The annotated English Gigaword corpus supported headline-generation experiments using article openings and headlines. CNN/Daily Mail pairs news stories with highlights and became a major benchmark for multi-sentence summarization. XSum, introduced by Shashi Narayan, Shay Cohen, and Mirella Lapata in 2018, pairs BBC articles with short summaries that often require substantial abstraction. Performance on these resources is not interchangeable: copying leading sentences is particularly competitive in some news settings but less suitable for highly compressed, rewritten outputs.

Long-document datasets expose other demands. Scientific-paper collections based on arXiv and PubMed pair papers with abstracts, which usually organize information around objectives, methods, results, and conclusions. BillSum, introduced in 2019, uses legislative documents and summaries. SAMSum provides conversational examples paired with summaries. Each resource makes assumptions about what counts as important: a paper abstract, a legislative synopsis, and a conversation digest encode different communicative purposes.

Standard supervised learning minimizes the negative log-likelihood of a reference summary:

L(θ)=−∑t=1Tlog⁡pθ(yt∗∣y<t∗,x).\mathcal{L}(\theta) =-\sum_{t=1}^{T}\log p_\theta(y_t^*\mid y_{<t}^*,x).

This objective teaches the model to predict a particular reference wording, even though many alternative summaries may be acceptable. Some systems supplement it with sequence-level rewards, preference comparisons, or reinforcement learning. These approaches can target human judgments or automatic metrics, but they also risk rewarding superficial properties if the optimization signal is incomplete.

Dataset construction determines how trustworthy an evaluation will be. Duplicate or near-duplicate stories can cross training and test boundaries, and coverage of the same event can create less obvious overlap. Such data leakage can inflate apparent generalization. Differences in publication dates, domains, languages, or document length create distribution shift. Reference summaries may also contain facts absent from the supplied input, making a faithful model appear deficient. Consequently, data provenance, input–reference alignment, and split design are substantive parts of summarization research rather than merely administrative details.

Long documents, multiple sources, and multilingual inputs

Long-document summarization must identify important information distributed across a source that may exceed a model’s context window. Simply truncating the input assumes that the beginning contains everything essential. That assumption can be tolerable for some news articles but is poorly suited to a report whose findings appear after extensive methods or background sections. Full self-attention also has a quadratic interaction pattern in sequence length, creating computational costs even when a long input technically fits.

One strategy is hierarchical summarization: divide the source into sections, summarize each section, and combine the intermediate outputs. This reduces the amount of text processed at any single stage, but early omissions can become permanent. Repeated compression may gradually strip qualifications, citations, or distinctions among entities. An alternative retrieves relevant passages before generation. Retrieval-augmented generation is especially useful for query-focused summaries, although retrieval must preserve enough surrounding context to interpret the selected evidence.

Multi-document summarization adds source alignment and conflict management. Reports may use different names for the same entity, repeat a shared wire-service account, or describe events at different stages. Counting repeated claims without considering shared provenance can mistake copying for independent corroboration. Temporal metadata can be crucial: an early estimate and a later correction are not necessarily competing accounts of the same evidential status. A faithful summary may need to preserve explicit disagreement rather than force all sources into a single narrative.

Conversations impose a different structure. Speakers interrupt one another, refer implicitly to shared objects, and revise proposals. A suggestion is not a decision, and an acknowledgment is not necessarily acceptance of an assigned task. Summaries therefore often track dialogue acts, participants, commitments, and unresolved questions. When the input comes from speech recognition, transcription mistakes and incorrect speaker attribution become additional sources of error. A polished meeting summary can conceal these upstream uncertainties unless the system retains evidence links.

Multilingual and cross-lingual summarization require both linguistic competence and appropriate content selection. Translating a source before summarizing can propagate translation errors; summarizing first can remove context needed for an accurate translation. Joint models avoid a rigid two-stage pipeline but still face limited training coverage for many languages. Evaluation must account for differences in morphology, segmentation, and acceptable paraphrase. Names, dates, measurements, and culturally specific institutions can require careful handling even when the surrounding prose is fluent, making surface similarity to a single reference an especially incomplete measure.

Evaluation and benchmark interpretation

Summary quality has several distinct dimensions: coverage of important content, relevance to the request, factual consistency, coherence, readability, concision, and compliance with output constraints. These properties can conflict. A longer output may cover more facts while violating a strict budget, and a highly compressed one may become ambiguous. Evaluation therefore needs an explicit task definition and should not treat a single score as a complete measure of usefulness.

ROUGE compares generated summaries with human-written references through overlapping units. ROUGE-N measures overlap of nn-grams, while ROUGE-L uses a longest common subsequence. For a simplified single-reference case, nn-gram recall can be written as

RROUGE-N=∑gmin⁡(cS(g),cR(g))∑gcR(g),R_{\text{ROUGE-N}} = \frac{\sum_g\min(c_S(g),c_R(g))} {\sum_g c_R(g)},

where cS(g)c_S(g) and cR(g)c_R(g) are the counts of nn-gram gg in the system and reference summaries. Implementations also report precision or F-scores and differ in preprocessing and aggregation. This makes specification of tokenization, stemming, and scoring settings important for reproducibility.

Overlap measures are inexpensive and useful for comparisons under stable conditions, but they have well-known limits. A correct paraphrase may share few words with a reference, while a sentence containing the wrong number may retain nearly all its wording. BERTScore uses contextual token representations to measure semantic similarity more flexibly. Other learned metrics estimate quality from training examples. Such methods broaden the evaluation signal but do not turn semantic resemblance into a reliable test of factual support.

Human evaluation can assess dimensions that automatic metrics miss. Annotators may score coherence, identify unsupported claims, or compare which of two outputs better serves a specified reader. The Pyramid method, developed by Ani Nenkova and Rebecca Passonneau, represents reference content as semantic units weighted by their occurrence across multiple human summaries. Task-based evaluation instead examines outcomes, such as whether readers can answer questions accurately or locate relevant documents more quickly. These approaches make the role of the reader more explicit than reference overlap alone.

Evaluation design also affects conclusions. Clear annotation guidelines, independent judgments, agreement analysis, and appropriate statistical testing help distinguish systematic improvements from noise. Model-based judges can scale assessment but may favor verbosity, familiar wording, or outputs resembling their own style. A benchmark result must therefore be interpreted alongside document genre, reference construction, length constraints, model access to source material, and the particular dimensions measured. A system ranked highly on news summarization may perform differently on technical reports or informal dialogue.

Factual consistency, uncertainty, and attribution

A summary is factually consistent when its claims are supported by the supplied evidence, with qualifications and relationships preserved. This is distinct from general real-world truth. A statement can be true but absent from the source, making it inappropriate in a source-bound summary. Conversely, accurately reporting that a document makes a disputed claim does not establish that claim as true. The distinction between reporting evidence and endorsing its content is fundamental to faithful summarization.

Errors include wrong entities, altered quantities, reversed relations, missing negation, incorrect timing, and unjustified certainty. A source saying that a committee “may consider a proposal” does not support a summary saying that it “approved the proposal.” A study reporting an association does not by itself establish causation. Numerical mistakes can be especially difficult to notice: a model may attach a correct percentage to the wrong population or confuse an absolute increase with a relative one.

Generative systems can produce hallucinations because their training rewards plausible continuations rather than imposing logical entailment from the source. Common patterns in training data may supply a detail that seems likely but is not documented. Extraction reduces some opportunities for invented wording, yet selective quotation and omitted context can still distort meaning. Fluency can make either kind of error harder to detect because readers often use grammatical confidence as an imperfect cue to reliability.

Automatic consistency checks use several approaches. Entailment-based systems test whether source passages support summary statements. Question-based methods derive questions from the summary, answer them using the source, and compare the answers. Other tools align named entities, dates, quantities, or proposition-level relations. Each approach has limitations: an entailment model may struggle with a long evidence chain, question generation may overlook a crucial qualifier, and entity matching does not establish that the relationship among entities is correct.

Evidence attribution makes checking more concrete by linking claims to supporting passages. A citation is useful only if the cited material supports the associated statement; the mere presence of a reference is not a fidelity guarantee. Provenance tracking can retain document versions, timestamps, passage identifiers, and the origins of intermediate summaries. Preserving uncertainty is equally important. Expressions such as “estimated,” “preliminary,” and “according to the report” often carry substantive information. Removing them can change a tentative observation into a categorical assertion without adding any obviously invented fact.

Applications, deployment, and continuing research problems

Summarization is deployed in search interfaces, news aggregation, enterprise document systems, email tools, and meeting software. It can provide a short preview before a reader opens a document or a structured account of a larger collection. In software development, summaries may describe changes, issue discussions, or technical documentation. In research workflows, they can help organize literature, although an abstract-like output does not replace examination of methods, evidence, and limitations.

Different applications require different failure tolerances and representations. A news preview may prioritize rapid orientation, whereas a contractual digest depends on exceptions, dates, and obligations. Scientific and clinical documents contain specialized distinctions that may disappear under aggressive compression. A summary of a record is also different from an independently validated professional judgment. System design therefore often includes explicit boundaries on the task, links to original passages, and human review where errors would have substantial consequences.

Deployment introduces privacy and security concerns beyond linguistic quality. Source documents may contain confidential information, and the resulting summary may concentrate sensitive facts into a more easily shared form. Permissions for accessing the original material must remain relevant to generated outputs. Systems that ingest untrusted documents also face prompt injection: malicious text may attempt to redirect the model away from summarization or induce disclosure of other information. Document content must be distinguished from instructions governing the application.

Operational constraints shape architecture. Long inputs increase processing costs and latency, while hierarchical pipelines add stages at which errors can accumulate. Caching can reduce repeated work but requires attention to document versions and access rights. Output schemas can make summaries easier to integrate into downstream software, although a valid structure does not establish correct content. Monitoring may track unsupported claims, omissions, length violations, and changes in performance when the document population shifts.

Continuing research focuses on preserving evidence through repeated transformations, evaluating usefulness without relying on a single reference, and controlling what a summary emphasizes. User-specific summaries raise questions about how to adapt relevance without suppressing important contrary evidence. Long-running collections require incremental summaries that incorporate corrections rather than merely append new statements. Another problem is determining when a requested degree of compression is incompatible with faithful coverage. Systems can satisfy a word limit while silently discarding essential qualifications; detecting that conflict requires a representation of which information is indispensable for the task, not merely a mechanism for producing shorter text.