Extractive Text Summarizer

Who this is for

It suits a specific job: you have a few paragraphs on a single topic — a report section, a long email, a set of meeting notes — and you want to see which sentences carry the repeated vocabulary before you write your own précis. Treat the output as three sentences worth reading first, not as a finished summary.

Because the result is copied verbatim, nothing can be invented. The risk runs the other way: a correct sentence can mislead once it is lifted away from the sentences around it. The sections below are about that risk.

How sentences are scored

1. Check the input. The text is trimmed. Empty text gives "Enter text to summarize." Text over 100,000 characters gives "Use up to 100,000 characters."
2. Split into sentences. The text is cut at every run of ., ! or ?, keeping the punctuation with the sentence before it. Each piece is trimmed and blanks are dropped. Text with no final punctuation is still one sentence. Text made only of .!? gives "Enter text containing words."
3. Reduce each sentence to content words. The sentence is lower-cased and broken into runs of letters or digits, in any alphabet. Tokens of 1 or 2 characters are dropped, and so are 30 common English function words — of which only these 15 are long enough to matter: the, and, but, for, with, are, was, were, this, that, from, you, they, our, your. "She", "her", "not" and "very" are not on the list and count as content words.
4. Count every content word across the whole text. One combined tally, not one per sentence. A word that occurs four times has the value 4 everywhere it appears.
5. Score each sentence. The score is the average value of its content words.
6. Rank and pick three. Highest score first; a tie goes to whichever sentence came first. A sentence identical to one already chosen is skipped. The top three are kept.
7. Restore the original order and join the three with a single space.

Because it is an average, a short sentence built almost entirely from the repeated topic word scores very highly, while a long sentence full of precise, once-only detail scores about 1 — detail is systematically penalised. A sentence with no content words at all is divided by 1, scoring 0, and comes last.

The scoring worked through

Input: Hello everyone. It is sunny. I like coffee. Solar panels make solar power. Solar power reduces fuel use. Solar panels need sunlight.

Across the whole text, solar occurs 4 times, panels 2 and power 2. Every other content word occurs once. "It", "is" and "I" are dropped as too short.

The top three are sentences 4, 6 and 5. Put back into document order they produce the output: Solar panels make solar power. Solar power reduces fuel use. Solar panels need sunlight. Note that "solar" appears twice in sentence 4 and is counted twice, in both the sum and the divisor, and that the three sentences tied at 1.00 are ranked by position.

Three ways the result breaks

A pronoun loses the name it referred to

Input: In 2019 a graduate called Fatima borrowed money from her uncle. She founded the bakery. The bakery supplies bread. The bakery bakes bread daily. The first sentence scores 1.00 — eight content words, all unique — so it is dropped, and the output begins: She founded the bakery. The summary opens with a pronoun whose referent is gone. Opening sentences introduce people in fresh words, so they are exactly the sentences this scoring throws away.

A list is announced and then half of it disappears

Input: Our refund policy has three steps. Open a ticket within 30 days. Attach the receipt. Wait for the refund. The refund takes five days. The output is Our refund policy has three steps. Wait for the refund. The refund takes five days. — it promises three steps and shows one of them. Short imperative items share almost no vocabulary, so procedures, recipes and bullet lists come out worse than continuous prose.

Dialogue is cut at the quotation marks

Input: Wait! he shouted. Did you see that? The truck stopped. The truck stopped just in time. The exclamation mark inside the speech ends a sentence, so "Wait!" and "he shouted." become two separate candidates, and the output is Wait! The truck stopped. The truck stopped just in time. The speaker has been separated from the speech. The same splitter also breaks Dr. Smith arrived. into two sentences and turns The rate is 3.5 percent. into "The rate is 3." and "5 percent."

How to check a summary before you use it

1. Read the three sentences on their own, as if you had never seen the source. Anything you understand only because you remember the original is a problem for your reader.
2. Point at the referent of every "she", "he", "it", "they", "this" and "the above". If it is not inside the summary, replace the pronoun with the name. Check numbers and dates the same way.
3. Search the source for negations and conditions — "unless", "only if", "does not", "except" — that sit in dropped sentences and reverse what the summary implies.
4. Check that any count the summary mentions ("three steps", "both options") is actually enumerated there.
5. If the output is almost as long as the input, the text had too few sentences for the selection to do anything. Summarise it yourself.

Limits

It always returns three sentences, or the whole text if there are fewer — the number is fixed in the code and cannot be changed from the page. It does not paraphrase, shorten a sentence, merge two sentences, generate a headline, pull out keywords, or produce bullet points. It cannot answer a question about the text.

It is tuned for continuous prose on one subject. Text covering several unrelated topics returns three sentences from whichever topic repeats its vocabulary most, and the rest vanish; headings and captions pasted in with the body compete as ordinary sentences. The word splitter handles any alphabet, but the stop-word list and the sentence punctuation are English, so scripts that do not end sentences with .!?, or that write no spaces between words, are not handled properly.

Questions

Reviewed October 1, 2026. Calculation methods and corrections