Extractive Summarization Explained: How a Text Summarizer Picks Sentences

An extractive summarizer makes a summary by choosing sentences from your text, not by writing new ones. It gives every sentence a score, keeps the highest-scoring ones, and puts them back in their original order. The score usually comes from three signals: how often the sentence's words appear in the whole text, where the sentence sits (the first and last sentences often matter more), and cue words such as "conclusion" or "important". An abstractive summarizer, the kind built on a language model, writes new sentences instead. Extractive summaries can be choppy, but every sentence in them is one you actually wrote, so they cannot invent a fact.

The FindUtils Text Summarizer is extractive and runs in your browser. This post explains where the method comes from, walks through a real eight-sentence example, and lists the traps that make a summary miss the point.

Where the Three Signals Come From

Word frequency (1958). H. P. Luhn at IBM proposed that the words an author repeats are the words the text is about. His paper, The Automatic Creation of Literature Abstracts (IBM Journal of Research and Development 2(2), 1958), scored sentences by how many of these frequent words they contain and picked the top ones. He also left out the most common words, such as "the" and "and", because they are frequent in every text.

Position and cue words (1969). H. P. Edmundson's New Methods in Automatic Extracting (Journal of the ACM 16(2), 1969) added more signals: words from the title and headings, the position of a sentence (first and last sentences of a paragraph or document), and "cue" words that often mark key sentences, such as "significant" or "in conclusion". He found that cue, title and position words did better than frequency alone.

Graphs (2004). TextRank (Mihalcea and Tarau, 2004) links each sentence to the sentences it shares words with, then ranks them with the same idea as Google's PageRank. A sentence that is similar to many others is treated as central. It needs no training data either.

Most quick summarizers on the web, including ours, use a mix of the first two ideas, because they are fast and easy to explain.

How the FindUtils Summarizer Scores a Sentence

1. Split the text into sentences at ., ! and ?.
2. Count words. Every word longer than three letters is counted across the whole text. Shorter words such as "the", "and" and "for" are ignored.
3. Score each sentence by adding up the counts of its words and dividing by the number of words, so long sentences do not win just by being long.
4. Apply position boosts. The first sentence is multiplied by 1.5 and the last by 1.2.
5. Apply cue words. The score is multiplied by 1.3 for each of these words found in the sentence: important, significant, conclusion, result, therefore, thus, key, main, primary, essential.
6. Keep the top sentences for the chosen length (Short is about 20% of the sentences, Medium about 40%, Long about 60%, or 10% to 80% with the slider) and show them in their original order.

A Worked Example

Here is an eight-sentence note about a bakery that wants online orders. With Medium (40%), the summarizer keeps round(8 × 0.4) = 3 sentences.

The summary reads: "A small bakery wants to sell bread online. An online order form would collect the address, the bread and the delivery slot in one place. The owner can start with one form and add payments later." That is a fair summary. Sentence 4 wins without any boost, because it uses the words the whole note repeats: "order", "form", "bread", "delivery", "online". Sentence 7, the one a person might pick as the point, comes fourth even with its cue word.

Five Traps That Make an Extractive Summary Miss the Point

1. Common long words still count. Words of four letters or more are counted, and the summarizer has no stopword list. "That", "with", "this", "have" and "from" can lift a sentence just because they are frequent. Luhn removed such words; a short-word filter only removes some of them.

2. Cue words are matched inside other words. The check looks for the cue word anywhere in the sentence, so "keyboard" matches "key", "domain" matches "main", and "results" matches "result". In a short test, "The keyboard shortcut list is long." got the 1.3 cue boost for "key" on top of its first-sentence boost, and "Our domain moved last month." got one for "main".

3. A last line without a full stop is dropped. Only text that ends in ., ! or ? counts as a sentence. If your last line is a heading-style phrase or has no full stop, it is not in the summary at all, and the last-sentence boost goes to the line before it. The same split also breaks "e.g." and "3.5" into extra sentences.

4. The length is in sentences, not words. "Short (~20%)" keeps a fifth of the sentences. If the kept sentences are long, the summary can be much more than 20% of your words. The Reduction figure shows the real cut in words.

5. No headings, no meaning. The method does not know what the text is about. It cannot merge two half-points into one sentence, and it will keep a sentence that starts with "This" even when the sentence it refers to was left out. Read the summary once and fix such links by hand.

Extractive or Abstractive?

Use an extractive summary when every sentence must be traceable to the source, such as meeting notes, research notes or a report you have to quote. Use an abstractive summary when a fluent overview matters more and you can check it against the original. You can also combine the two: pick the key sentences first, then have a model rewrite only those, so the model works from a short, checked extract.

Getting Better Summaries

• Clean the text first. Remove menus, footers and repeated headings, and make sure each line ends with a full stop. For a PDF, extract the text with the PDF Text Extractor and clean it before you summarize.
• Start with Medium, then go shorter if the result still repeats itself.
• Put the point in a plain sentence. A sentence that uses the text's main words and says what matters is the one the summarizer picks, and also the one your readers remember.
• Check readability after. The kept sentences are often the longest ones. The Readability Calculator shows whether the summary is harder to read than the original; our Readability Calculator guide explains the scores.

Need help with a project? The team behind FindUtils takes on freelance work through Upwork, including adding an AI feature to your web app, such as a chat assistant, document analysis or search.

FAQ

What is the difference between extractive and abstractive summarization?

Extractive summarization copies the most important sentences from the text. Abstractive summarization writes new sentences that say the same thing in fewer words, which reads better but can introduce errors.

Why did the summarizer skip the sentence I think is most important?

Its score depends on repeated words, position and cue words, not meaning. A key sentence that uses different words from the rest of the text can score low. Rewrite it with the text's main terms, or pick a longer summary.

Does extractive summarization use AI?

Not in the sense of a language model. It is a scoring method from the 1950s and 1960s. Graph methods like TextRank are also unsupervised and need no training data.

Can I summarize text in other languages?

The method works on any language that separates words with spaces and ends sentences with ., ! or ?, but cue words are English, and accented letters are removed before words are counted, so results are best in English.

Is my text uploaded?

No. On this page, the summary is made in your browser.

Related Reading

• Readability Calculator guide: check the reading level of any text
• Readability scores explained: why Flesch, Fog and SMOG disagree
• Word Counter guide

Sources

• Luhn, H. P. (1958): The Automatic Creation of Literature Abstracts, IBM Journal of Research and Development 2(2), 159-165
• Edmundson, H. P. (1969): New Methods in Automatic Extracting, Journal of the ACM 16(2), 264-285
• Mihalcea, R. and Tarau, P. (2004): TextRank: Bringing Order into Text, EMNLP 2004
• Wikipedia: Automatic summarization