Article in, one post out Free · No sign-up · Runs in your browser

Guide 06

Extractive vs Abstractive Summarization: What's the Difference?

One copies sentences, the other writes new ones. What that means for accuracy and readability.

Short answer

An extractive summary picks and trims sentences that already appear in the text, so it can't invent facts but may read choppily; an abstractive summary writes new sentences, so it reads smoothly but can introduce errors the source never contained.

Key takeaways

  • Extractive methods select sentences from the source; abstractive methods generate new wording.
  • Extractive output is easy to check because every sentence can be found in the original.
  • Abstractive output reads better but can state things the source doesn't say.
  • Either way, read the summary against the source before you post it.

Summaries come in two basic kinds. An extractive summary is made of pieces of the original: whole sentences or clauses, selected and trimmed. An abstractive summary is written fresh, in new words, the way a person would explain an article to a friend. The difference sounds technical, but it decides what can go wrong with a summary and how you should check it before posting.

What is extractive summarization?

Extractive summarization builds a summary by selecting the most important sentences from the source and leaving the rest out.

Think of a highlighter. You read an article, mark the three sentences that carry the main point, and copy them out in order. Software does the same by scoring every sentence. Common signals include:

  • Word frequency. Sentences that contain the words the article uses most are probably about its main topic.
  • Position. News stories put the key facts near the top, so early sentences score higher.
  • Length. Very short sentences rarely carry a full point; very long ones are hard to fit.

The highest-scoring sentences are kept, sometimes trimmed, and joined. The TweetSum summarizer works this way. It picks sentences for a post of 280 characters or less, a few key points and a one-line TL;DR, and every word in the output comes from your text.

What is abstractive summarization?

Abstractive summarization writes new sentences that express the main ideas of the source in different words.

This is how people summarize. You read the article, close it, and say what it was about. Software that does this, including the large language models behind most AI chat tools, generates fresh text based on what it has read. The result can merge several sentences into one, simplify jargon and change the order to make a clearer point.

Where does each one go wrong?

Extractive summaries go wrong by losing context; abstractive summaries go wrong by adding things that aren't there.

Extractive Abstractive
Can it invent a fact? No Yes
Can it drop a key qualifier? Yes Yes
Reads smoothly? Often not Usually
Easy to check against the source? Yes, find each sentence Harder, check each claim
Typical failure "This" or "But" with nothing before it A number, name or cause the source never gave

The extractive failure is easy to spot. A sentence that began "However, this figure fell" reads oddly on its own, and you notice. The abstractive failure is harder to spot, because the made-up claim reads just as fluently as the true ones. Researchers call these errors hallucinations or unfaithful summaries, and they are the main reason abstractive output needs a careful read.

Which should you use for a post on X?

Use extractive output as a draft you can trust word for word, then edit it lightly for flow.

For a post that shares an article, accuracy matters more than polish. Your followers will treat the post as a claim about what the article says. An extractive draft gives you sentences you know are in the source. You then:

  1. Fix any dangling openers ("This," "But," "He") by naming what they refer to.
  2. Cut words until it fits, keeping numbers and qualifiers.
  3. Add the link and credit the source.

Our guide to summarizing an article in one post walks through those steps in detail.

When is abstractive summarizing the better choice?

When the source is long, badly organized or full of jargon, and you're prepared to check every claim.

Some texts don't have three good sentences to extract: a meeting transcript, a rambling interview, a dense legal document. There, a rewritten summary is more useful. Write it yourself or use a tool, but then go back through the source and confirm each statement, especially numbers, names and cause-and-effect claims. Our guide on avoiding misleading summaries lists the errors to look for.

Do human writers use both?

Yes; most good summaries start extractive and end abstractive.

A careful writer underlines the key sentences first, then rewrites them in plain words, checking each one against the original as they go. The underlining step keeps the summary faithful. The rewriting step makes it readable. Purdue OWL's writing guidance describes a summary as the main ideas in your own words, much shorter than the source, and credited to it. That definition fits this two-step method well.

Questions

Is ChatGPT-style summarizing extractive or abstractive?

Large language models summarize abstractively: they write new sentences based on the text. They can also quote, but the default output is new wording, which is why it should be checked against the source.

Is TweetSum extractive or abstractive?

Extractive. It scores the sentences already in your text by word frequency, position and length, then picks and trims the best ones. It never writes new sentences.

Which kind of summary is more accurate?

Extractive summaries can't add facts that aren't in the text, but they can still mislead by leaving out a key qualifier or pulling a sentence out of context. Abstractive summaries can do both of those and also introduce new errors.

Can I combine the two?

Yes, and most good human summaries do. Start from the sentences that carry the main facts, then rewrite them in your own words, checking each claim against the source.

Sources