Text summarization is the natural language processing task of producing a shorter version of a text that preserves its essential information, either by selecting the most important sentences or by generating new wording. In speech & language AI, it is applied to articles, reports, meeting transcripts, support conversations and legal or medical records, wherever people need the gist of more text than they can read.
Two families dominate: extractive summarization, which copies existing sentences, and abstractive summarization, which writes a new text with a sequence-to-sequence or large language model. The central difficulty is the same for both: deciding what matters and staying faithful to the source.
How extractive and abstractive summarization work
Extractive methods score each sentence of the source and assemble the highest-scoring ones into the summary. Classic scoring uses word frequency, sentence position and graph centrality, as in the TextRank and LexRank algorithms; newer systems use a neural encoder that predicts whether each sentence belongs in the summary. Because every sentence is copied verbatim, extractive output is grammatical and traceable, but it can be choppy and redundant.
Abstractive methods read the full input and generate a new text token by token. Encoder-decoder transformers trained on document-summary pairs and large language models guided by prompts both belong here. They can paraphrase, merge facts from distant sentences and adapt tone, but they can also produce hallucinated details that the source never stated.
| Property | Extractive | Abstractive |
|---|---|---|
| Output wording | Copied source sentences | Newly generated text |
| Fluency | Often disjointed | Usually smooth |
| Faithfulness risk | Low, but context can be lost | Hallucination is possible |
| Compression flexibility | Limited by sentence length | Can reach any target length |
| Compute cost | Low | Moderate to high |
| Traceability | Every sentence has a source | Requires separate attribution |
Many production systems are hybrid: they first extract relevant passages to fit the model’s context window and then generate an abstractive summary from those passages, which reduces both cost and the room for invention.
Summary types and key parameters
Summarization is configured along several dimensions that change the task considerably. The compression ratio, the ratio of summary length to source length, is the most visible one; news summaries often keep roughly 5 to 15 percent of the words, while executive briefs of long reports may keep far less.
| Dimension | Options | Effect on the result |
|---|---|---|
| Input scope | Single document or multi-document | Multi-document needs deduplication |
| Purpose | Generic or query-focused | Query-focused keeps only relevant facts |
| Form | Headline, bullet list, paragraph | Shapes readability and length |
| Compression ratio | From a single line to about a third | Tighter ratios drop more detail |
| Input type | Written text or speech transcript | Transcripts need disfluency handling |
| Audience | Expert or general reader | Changes terminology and depth |
Long-document summarization adds a structural problem: inputs larger than the model’s context window are split into chunks, summarized separately and then combined, a pattern known as map-reduce or hierarchical summarization. Each merge step can lose details, so chunk boundaries should follow sections rather than fixed token counts.
With large language models, these parameters are set mostly through the prompt and decoding settings. Effective instructions name the target length in words or bullets, the audience, the fields that must be preserved and the rule to use only information from the source. A low sampling temperature reduces variation between runs, which matters when summaries are compared or audited, and a fixed output template makes results easier to check automatically.
How summarization relates to neighboring language tasks
Text summarization belongs to natural language generation but depends heavily on natural language understanding, because the system must identify salient content before rewriting it. It is often chained with speech-to-text for meeting summaries and with named entity recognition to guarantee that names, dates and amounts survive compression.
| Task | Input | Output | Key difference |
|---|---|---|---|
| Text summarization | Long text | Short text with main points | Reduces length, keeps meaning |
| Question answering | Text plus a question | Specific answer | Targets one information need |
| Text classification | Text | Category label | No text is generated |
| Translation | Text in one language | Same content in another | Keeps length and detail |
| Keyword extraction | Text | List of terms | No sentences or relations |
Evaluating summary quality
Summary quality has at least four dimensions: coverage of key content, factual consistency, coherence and conciseness. Automatic metrics capture only parts of this, so serious evaluation combines metrics with human review against the source.
| Method | What it measures | Main weakness |
|---|---|---|
| ROUGE | N-gram overlap with reference summaries | Rewards copying, ignores facts |
| BERTScore | Semantic similarity of embeddings | Still blind to factual errors |
| Entailment-based checks | Whether the source supports each claim | Struggles with long inputs |
| Question-based checks | Answers from summary versus source | Depends on generated questions |
| Human rating | Faithfulness, coverage, fluency | Slow and costly |
| Model-as-judge | Rubric scores from another model | Shares biases of language models |
A high ROUGE score does not prove that a summary is correct. Abstractive systems can score well while swapping a number or a negation, so factual checks are mandatory for regulated or customer-facing content.
Limitations and practical safeguards
The main risks of automatic summaries are well understood, and each has a practical countermeasure.
- Hallucination: generated facts not present in the source; mitigate with grounding prompts, source quotes and entailment checks.
- Omission of critical details: a short summary may drop a deadline or a dosage; define must-keep fields such as dates, amounts and obligations.
- Position bias: models overweight the beginning and end of long inputs; chunk by section and summarize each part.
- Lost hedging: phrases such as may, unless or except disappear, turning uncertain statements into firm claims.
- Transcript noise: speech recognition errors in names and numbers propagate into the summary unless corrected first.
For high-stakes use, the safest pattern is summaries with traceable citations to source passages, so that a reviewer can verify each statement in seconds instead of rereading the whole document.
In simple terms
Text summarization is what a good assistant does when handed a fifty-page report and asked for one page. Some assistants underline the most important sentences and copy them out. Others read everything and write a short note in their own words.
Computers do both. The copying approach is safe but sometimes awkward; the rewriting approach reads better but can occasionally invent things. That is why a summary used for important decisions should always be checked against the original.
A practical example
A customer support center records around 4,000 calls a day. Each call is converted to text with speech-to-text, and a language model then produces a five-line summary with fixed fields: customer issue, steps already taken, promised follow-up and deadline. The prompt instructs the model to use only information stated in the transcript and to quote the customer’s exact order number.
A rule-based check compares order numbers, dates and amounts in the summary with the transcript and flags mismatches for an agent to correct. Agents stop writing notes by hand, saving several minutes per call, and supervisors can review a whole shift’s issues in one sitting. The combination of generation and automatic fact checks keeps the summaries short while preventing the most damaging errors.