Text summariser

Pull out the key sentences of a long text, locally. Free, no upload. 1 GB max · Up to 1 GB · Processed locally, never sent to a server.

A board pack that lands the night before the meeting, a support thread that has been growing for three weeks, a policy document someone forwarded with no explanation: in each case you want the skeleton of a text before committing to it. This page pulls that skeleton out in a fraction of a second, and it is worth saying straight away how. It does not write a summary, it picks one. Every sentence displayed is yours, reproduced word for word and left in its original order.

The ranking is built on vocabulary alone. The text is cut into sentences at full stops, exclamation marks, question marks and ellipses. Every word of three letters or more is counted, apart from a closed list of 114 function words. A sentence is then worth the average frequency of its content words, measured against the most frequent word in the whole text. Three corrections are applied on top: the first sentence gains 35%, the last gains 10%, and any sentence shorter than 35 characters drops to 60% of its score.

No language model is involved anywhere in that chain, whatever the category heading above the page suggests. That is a constraint, and it is also the main guarantee the method offers: a summary that produces no new sentence cannot invent a figure, attribute a statement to the wrong person, or reach a conclusion on your behalf.

How to use it

  1. Paste at least three sentences Below that threshold a "Text too short" message appears and nothing is calculated. Ranking has no meaning when there is nothing left to leave out.
  2. Set the summary length Short keeps about 15% of the sentences, Medium 30%, Detailed 45%. The proportion applies to the number of sentences, never to the number of words, with a floor of two sentences.
  3. Press Run The heading above the result states the ratio you obtained, in the form "Summary (4 key sentences out of 12)". The calculation is instant and triggers no network request.
  4. Read the joins The selected sentences come back in their original order, but detached from their context. Check that no "it", "this rise" or "the latter" points at a passage that was left out.
  5. Pick how you copy One copy button strings the sentences together as a single continuous paragraph. The other puts each sentence on its own line, preceded by a bullet.

How a sentence earns its score

Counting words comes before everything else. Each form is recorded exactly as written, with no lemmatization: "decision" and "decisions" stay two separate entries, which spreads one theme across two counters. A sentence score is the sum of its content words frequencies divided by how many there are, with a denominator that never falls below six. That floor stops a three-word sentence from earning a flattering average simply because it happens to contain one frequent term.

The consequence is clear. The method works best on a text that insists. A report that comes back to the word "budget" ten times announces its own subject, and the sentences carrying that word rise to the top. A narrative, a letter, or a chain of arguments with no recurring term produce scores bunched tightly together, and the selection ends up being decided by the position bonuses rather than by the content.

What extraction guarantees, and what it cannot do

An extractive summary writes nothing. It cannot get an amount wrong, put a sentence in the wrong speaker mouth, or add a qualification that was absent from the source. The sentences you get are quotable as they stand, and any reader can find them in the original with a plain text search. That is precisely what a generative model cannot promise.

The other side is just as clear. No transition is manufactured between two sentences six paragraphs apart, so the reading is abrupt. Two complementary half-sentences will never be merged. An idea carried by three consecutive sentences, each unremarkable on its own, will be missed entirely. A language model would return a shorter, smoother text at equal information, at the cost of checking it sentence by sentence against the source. For sorting a pile of documents or spotting the passages worth reading, selection is enough. For a text meant to be published, it is raw material, not a result.

The texts where the method goes off the rails

Sentence splitting is mechanical and knows no abbreviations, and English pays a higher price here than French does. "Mr. Smith signed." is counted as two fragments, and "Mr." is three characters long, just above the two-character floor that discards debris. It therefore enters the sentence count as a fragment scoring zero, and inflates the total the percentage is applied to. "Inc." and "Fig." behave the same way.

Decimal points split sentences too. "The rate rose to 3.5 percent." breaks in two at the point, and both halves survive and are scored separately. Since English writes decimals with a point rather than a comma, any text carrying figures runs into this repeatedly.

Two further profiles resist the method. A bullet list pasted with no closing punctuation forms one giant sentence, which comes out whole or not at all; a recording transcript produces the same single block. Finally, the function word list is predominantly French. Around thirty common English words are appended to it, so "the", "and", "that", "with", "from" and "these" are filtered out, while frequent words such as "have", "will", "their", "which", "would" and "about" are still counted as content words and skew the frequencies. A Spanish, German or Italian text gets no filtering at all, and its determiners dominate the ranking.

What the site keeps of your text

The whole treatment amounts to a handful of regular expressions and two sorts, executed in the open tab. There is no service to query, so there is no text to transmit, and the page keeps working if your connection drops after it has loaded. The sidebar keeps a trace of the operation, limited to the word "Summary" and the count "12 to 4 sentences": neither the text you pasted nor the sentences that were selected appear in it. That trace lives in your browser local storage, is capped at four entries, and the clear link empties it.

Frequently asked questions

Why is my first sentence almost always kept?

It receives a 35% bonus, and the last sentence a 10% one. The setting targets the usual construction of an article or a report, where the essential point arrives immediately. If your text opens with a greeting or a pleasantry, delete it before pasting.

Can I get three lines out of ten pages?

No. The proportion applies to the sentence count, so two hundred sentences reduced to 15% still leave thirty. For a very long document, treat it section by section, then run the assembled result through the tool a second time.

Does the result change if I run it again?

Never. No random draw is involved at any point. With identical text and identical settings the selection is rigorously the same, today and in six months.

Can a short but decisive sentence be dropped?

Yes, and that is the main blind spot. "The contract is terminated." is penalized twice over: its score is divided by six words at minimum, then cut by 40% because it runs to fewer than 35 characters.

Does the summary take headings into account?

No, they are treated as ordinary text. A subheading with no closing punctuation is merged into the sentence that follows it, and its short length counts against it in the ranking.

Can I summarize a PDF or a Word document?

Not directly, the field only accepts text. Convert the document first, or run optical recognition on it if it is a scan, then paste the result here. Formatting is ignored in any case.

Similar tools

Discover other tools

Categories