Chapter 04 · Section I · 10 min read
4.1 Summarization: extractive vs abstractive
The two flavours of summarization, when each is right, and the prompt shapes that make each one behave. Worked with a Nepali news article.
“Summarize this” is one of the most common prompts, and one of the most under-specified. There are two very different jobs hiding inside that verb. Once you can name which one you want, the prompt writes itself.
The two flavours
Extractive summarization — pick sentences (or fragments) from the source and stitch them together. Every word in the summary appears in the source. Faithful, verifiable, often clunky.
Abstractive summarization — write a new summary in your own words that captures the source’s meaning. More readable, more likely to introduce distortion.
Almost every summarization complaint (“it left out the important part,” “it made stuff up,” “it’s a paraphrase, not the actual news”) is really a mismatch between what the user wanted and which flavour the model produced.
Extractive — the prompt
Extract the three sentences from the article that best convey the news.
Rules:
- Copy sentences verbatim; do not paraphrase.
- Preserve their original order.
- Do not add any connecting words or transitions.
<article>
The Kathmandu Metropolitan City today announced a NPR 240 million
budget to widen the Ring Road from four lanes to six between Kalanki
and Balaju. Mayor Balen Shah said work would begin in early Kartik
and finish before the next monsoon. Local shopkeepers along the
affected stretch expressed concern about temporary business loss
during construction. The city said compensation guidelines would be
published within two weeks.
</article>Extractive is right when:
- Accuracy matters more than style (legal, medical, official announcements).
- The reader will scan for specific facts.
- You need to preserve source language for later citation.
Abstractive — the prompt
Write a 60-word news brief summarizing the article below, in your own
words, in the voice of a plain-English morning newsletter for busy
professionals in Kathmandu.
Include: the decision, the budget, the timeline, and one concrete
concern raised.
<article>
[same article]
</article>Abstractive is right when:
- The reader wants a fast digest, not a citation.
- The source is long and needs compression.
- Register matters (email brief, headline, meeting note).
Common failure modes
Left out the lede. The model buried the main news in the middle. Fix: specify what must appear. “Include the decision, the budget, and the timeline.”
Made up a number. The model summarized “over 200 million rupees” as “300 million rupees.” Fix: switch to extractive for numeric facts, or add “Do not restate numbers that are not exactly present in the source.”
Turned a specific into a generic. “Mayor Balen Shah announced” became “the government announced.” Fix: “Preserve every named person, place, and organisation exactly as written.”
One long paragraph when you wanted bullets. A format issue, not a summarization issue. Fix: “Return three bullets, no preamble.”
The dial: length, audience, focus
Every good summarization prompt sets three things explicitly:
- Length — words, sentences, or bullets. “Short” is not a length.
- Audience — busy PM, high-school newspaper reader, engineer, minister’s chief of staff. Same content, different tone.
- Focus — what must appear. Without focus, the model picks whatever is textually salient, which is often the intro paragraph rather than the actual news.
The hybrid pattern
For high-stakes work, run both:
For each key claim in the article, return:
1. A one-line abstractive summary in your own words.
2. The verbatim sentence(s) from the source that support it.
Format as Markdown, one claim per section.The abstractive line gives you readability. The extractive quote gives you an audit trail. This is the pattern used by the best news-summary tools and by anyone who has been burned by a made-up statistic in a summary.
Check your understanding
Quick check
—