Extractive summarization builds a summary by picking sentences out of the source text and putting them together as they are. No new sentence is written. The system splits the text into units, scores each one for importance, and lines up the highest scorers in their original order.
Scoring can rest on several signals. Classic methods look at word frequency, where the sentence sits in the document, and overlap with the title. Graph-based approaches connect sentences by similarity and pick the ones nearest the centre. Embedding methods surface the sentences closest to the average vector representing the document.
Its counterpart is abstractive summarization, where the model writes its own sentences. That reads more smoothly and carries more risk. The advantage of the extractive route sits right there: because every sentence comes from the source, nothing gets invented and each statement can be checked against the original word for word.
A concrete case: a legal team looking for termination conditions in a 40-page contract sees the relevant clauses in the contract's own wording. An abstractive summary doing the same job might shift the legal meaning while rephrasing.
The weakness is flow. With no connective tissue between the selected sentences the summary can read as disjointed, and pronouns inside those sentences lose the context they referred to.