AI oppsummering

Extractive methods select existing passages; abstractive methods generate new wording. A good summary preserves the important claims, uncertainty, and relationships without adding facts that the source does not establish.

Viktige takeaways

• Define the summary’s purpose.
• Check omissions separately from invented additions.
• Keep links to the original evidence.

Define what the summary is for. A decision brief, a meeting record, and an introductory overview should preserve different details. Specify the audience, length, and required fields before judging the output. Shorter is not automatically better if it removes essential conditions. Check coverage and factual consistency separately. A summary can accurately describe one section while omitting the central result. It can also cover every topic while attributing a statement to the wrong person or reversing a qualification. Long documents may need staged processing, but summaries of summaries can lose context. Preserve links to original passages and retain the distinction between confirmed results, proposals, and commentary. For documents with tables, verify units, row labels, and footnotes instead of treating extracted text as a complete representation. Evaluate on realistic documents and score error types explicitly. Unsupported claims, omitted decisions, incorrect quantities, and changed dates deserve separate attention. Give readers a route back to the source, and label any additional interpretation as analysis rather than part of the summary.

1. In an invented meeting note, one participant proposes a Friday release. The group decides to run another test before selecting a date.
2. A draft summary says “The team approved a Friday release.” Mark this as an unsupported decision.
3. Replace it with “Friday was proposed; the date remains undecided pending testing,” and retain the source passage.

This exercise checks whether the summary preserves the actual status of a claim.