Is the summary written by the tool or taken from my text?

Taken from your text. The summarizer is extractive: it selects whole sentences and returns them verbatim in their original order. It never composes, shortens or paraphrases, which is why the output can read as disjointed and is always labelled extractive.

What do the three scores in the sentence table mean?

Centrality is TextRank PageRank over a sentence-similarity graph, frequency is the summed log weight of the sentence's content words divided by the square root of its length, and position is a mild decay from the start of the text. The final score weights them 0.55, 0.35 and 0.10.

Why did the summarizer pick a sentence I would not have chosen?

Because the scoring rewards vocabulary shared with the rest of the text, not importance. A sentence that repeats the document's key terms ranks above one carrying a decisive but unusual claim. The score table shows which signal raised it so you can disagree with the reason, not just the result.

How is key-term coverage calculated?

The tool takes the ten content terms with the highest weight, where a term's weight is 1 + log(count), and reports what share of their combined weight appears in the selected sentences. It is a measure of what the summary kept, not a quality judgement.

Does the same text always produce the same summary?

Yes. PageRank starts from a uniform vector and runs a fixed 30 iterations with deterministic arithmetic, so there is no randomness and no model sampling. The same input produces the same output on every run, in every browser.

Text Summarizer

Your text

Paste the text to summarise

Summary length

The percentage is applied to the sentence count and rounded to at least one sentence.

Extractive summary

Start typing above — the results update as you write.

How the scoring works

Each sentence scores 0.55 × TextRank centrality + 0.35 × log term frequency (divided by the square root of the sentence length) + 0.10 × position decay, normalised so the best sentence is 1.000. Centrality uses the classic shared-content-word similarity |Si ∩ Sj| / (log |Si| + log |Sj|) and 30 PageRank iterations; the three parts are shown in the table so any ranking can be recomputed by hand.

This is a heuristic, not an understanding of your document. It rewards sentences that share terms with the rest of the text, so an introduction that repeats the topic can outrank a conclusion that matters more; a sentence with unusual vocabulary may be underrated. Short texts and texts with one dominant term are where it is least reliable. Method: TextRank centrality (0.55) + log term frequency per sentence (0.35) + position (0.10), all normalised to the best sentence.

This text summarizer scores every sentence with TextRank centrality, log term frequency and position, then returns the highest-ranked sentences in their original order as an extractive summary — the sentences are taken from your text, never rewritten.

What is Text Summarizer?

Most summarisers on the web either call a model or quietly paraphrase. This text summarizer does neither. It was built for one job: summarize text by selecting the sentences that already say what it says, so every sentence in the output already exists in the input, in the same words and the same order. Nothing is composed, shortened or reworded, and the page labels the result an extractive summary for exactly that reason. The engine runs in the browser: no model, no training data, no network request, and nothing is uploaded.

Sentence scoring has three published parts. TextRank centrality is the first: sentences become vectors of their content words, a similarity graph is built with the classic Mihalcea-Tarau measure |Si shared with Sj| / (log |Si| + log |Sj|), and PageRank runs for 30 fixed iterations with damping 0.85. Log term frequency is the second: each content word weighs 1 + log(count), summed over the sentence and divided by the square root of its length so long sentences do not win by size alone. Position is the third: a mild 1 / (1 + 0.12 x index) decay. The final score is 0.55 x centrality + 0.35 x frequency + 0.10 x position, normalised so the best sentence reads 1.000. Every part is printed per sentence in the score table, so a ranking can be recomputed rather than trusted.

Length is a control, not a guess. Choosing about 10, 20, 30 or 40 percent sets the selection size to that share of the sentence count, rounded to at least one sentence. The tool reports how many sentences were selected, what share of the words that is, and the key-term coverage: the top ten content terms by weight, and the share of their total weight the selection contains. Coverage is the honest measure of what a summary kept, and it is more informative than a length percentage on its own.

Where the method is weak, the page says so. Repetition beats meaning: a sentence that shares vocabulary with the rest of the text outranks one that carries the single most important claim but uses unusual words. Texts with one dominant term, very short texts, and documents whose opening paragraph restates the title are the cases where the ranking is least useful. The sentence splitter is also a mechanical rule, so headings, captions and list items are treated as sentences, abbreviations such as Dr. can split a sentence in two, and negation is invisible — a summary can quote a sentence that says the opposite of the document's point if the vocabulary matches.

In practice the tool is a triage aid, not an author. Run it at 10 percent to see what the scoring thinks the text is about, then at 30 percent for a working brief, then read the original for anything the ranking missed by design. Because the output is a quotation, the seams between selected sentences show; that is information about the document, not a bug to smooth over.

The result is deterministic. The PageRank starts from a uniform vector, performs a fixed number of iterations and involves no randomness, so the same text produces the same ranking every time. Your text is processed in this browser only and is not stored; the summary disappears with the tab unless you copy or download it.

What this is not: an abstract generator, a rewriter, a translation tool or a fact checker. It cannot judge whether the extracted sentences are true or whether the document argues well, and it does not claim to understand the text — it ranks sentences by three stated signals and shows its arithmetic.

How to use Text Summarizer

Paste the text you want to summarise; the sentence scores appear as you type.
Choose a length — 10 percent for a triage read, 30 percent for a fuller brief.
Read the summary, then check the key-term coverage figure to see what the selection kept.
Open the score table and look at the three parts of any sentence you think was misjudged.
Copy or download the report, and edit the selection by hand where the scoring missed your point.

When to use Text Summarizer vs related tools

Use this summarizer when you need a shorter version of a document you already have and want to know exactly which sentences survived. When you need the result as a ranked list rather than a paragraph, the Key Points Extractor shows the same scoring as numbered points with their evidence, and when you need a single line that fits a strict character budget, the TL;DR Generator is built for that.

If the question is how the text reads rather than what it says, the Readability Grade Scorer measures difficulty and the Sentence Variety Checker looks at rhythm.

Privacy & Security

This tool runs entirely in your browser — no data ever leaves your device. There is no server round-trip, no upload, no logging, and no account required. Your input is processed locally using client-side JavaScript and is never stored, transmitted, or accessible to anyone else. When you close the tab, everything disappears.