Skip to content

What an extractive summary can and cannot tell you

There are two kinds of automatic summary. One rewrites a document in new words, which needs a language model and can state things the document never said. The other picks the sentences that carry most of the document's own vocabulary and lists them as they are. This tool does the second, so every line it gives you is in the PDF, on the page it names.

Updated

Which tool you need

  • A report to skim before a meeting

    Choose Short for the gist, then follow the page numbers to the sections that matter and read those in full.

  • A paper or a long contract

    Medium or Long gives a fuller picture. Read the key terms underneath as well: they show what the document keeps coming back to.

  • Only one chapter or section matters

    Choose just its pages before summarizing, so sentences from the rest cannot crowd it out.

  • You want every word rather than a selection

    PDF to Text saves the whole text of the pages you choose.

    PDF to Text

Why nothing is rewritten

A rewritten summary reads more smoothly, and it can also be confidently wrong: a model can merge two figures, drop a 'not' or invent a conclusion, and nothing on the page tells you which. A list of the document's own sentences cannot do that, and each one comes with a page to check it against.

The cost is fluency. A picked sentence can open with 'this' or 'however' and point back at something left out, which is why the page number matters: it takes you straight to the context.

What the summary file looks like

The .txt opens with the file's name and how many of the document's sentences were kept. The key points follow, one to a line, each ending with the page it came from in brackets, and the last line lists up to eight key terms.

Headings are never picked: they mark where sections begin, and a section's opening sentences count for a little more. Fragments under four words, rows of figures and short lines that repeat from page to page, like a running header, are passed over as layout rather than content.

When the summary comes out thin

The length you choose is a target, not a promise. No summary keeps more than about half of the sentences on the pages you chose, so a short document can give fewer than Long asks for. A sentence too close to one already chosen is skipped, so the space goes to something the summary does not yet say.

Pages of headings, labels and figures may hold no sentence worth picking at all. Then the result says there is nothing to pick, rather than offering a summary made of captions.

What to know before you start

  • Sentences are picked whole; none is shortened, merged or reworded.
  • Prose in English summarizes best, and slides or forms summarize badly.
  • Scanned pages have no text to pick from.
  • Text set sideways or vertically on a page, such as a margin note, is not considered.

Read next