Summarisation Shortening a document, and knowing whether the short version is still true.
Short answer. Summarisation condenses a text into a shorter version that keeps the main information. There are two broad approaches: extractive methods build the summary only from passages copied out of the source, while abstractive methods may write new words and phrases, as a human-written abstract does. Language models mostly write abstractive summaries, which makes them fluent and also prone to adding things the source never said: a 2020 human evaluation found hallucinated content in more than 70% of the single-sentence summaries it checked. For an office, the safeguard is to keep each line of a summary tied to its source passage.
01
Extractive and abstractive
The distinction comes from the summarisation literature. Stanford and Google researchers put it this way: extractive methods assemble summaries exclusively from passages, usually whole sentences, taken directly from the source text, while abstractive methods may generate novel words and phrases not featured in the source, as a human-written abstract usually does. They note the extractive approach is easier.
Approach
How it works
Strength
Risk
Extractive
Selects whole sentences or passages from the source
Wording is the source’s own
Can read as a choppy list and miss the point
Abstractive
Writes new sentences that restate the source
Fluent; can combine ideas
Can distort or add facts
02
Where abstractive summaries go wrong
A 2020 Google study had trained annotators check one-sentence summaries of 500 BBC news articles from the XSum dataset, written by several neural systems and by the dataset’s journalists. It separated two kinds of hallucination. Intrinsic ones use terms from the document but misrepresent it. Extrinsic ones add information that is not in the document at all. Together they appeared in more than 70% of the single-sentence summaries, and over 90% of the extrinsic additions were wrong. Models that started from pretrained parameters did best.
Those were 2020 systems on a short-summary task, so newer models may do better, but the lesson that carries over is that fluency is not accuracy. A second finding applies to long documents: researchers at Stanford and elsewhere found that language models often use information at the beginning or end of a long input best, and do worse when the relevant passage is in the middle, even for models built for long contexts.
Failure
Example in an office
How to catch it
Misstated detail (intrinsic)
A fee of ₹4.2 lakh becomes ₹42 lakh
Check every number and date against the source
Invented content (extrinsic)
A clause or decision that is not in the document
Require a quote or page reference for each claim
Missing middle
An exception on page 40 of a 100-page tender is left out
Summarise by section, and spot-check the middle pages
Wrong attribution
A task given to the wrong person in a meeting summary
Link each item to its transcript line and confirm owners
03
Making a summary checkable
Tie every sentence to the passage it came from, with a page or timestamp.
Keep the exact words for numbers, dates, names and clauses; paraphrase only the narrative.
Summarise long documents section by section and show the sections, so nothing in the middle is skipped silently.
Tell the model to say “not in the document” when the answer is missing, and test that it does.
Keep the original one click away, and have a person read the part that matters before acting.
Do not rely on a summary when the exact wording is the point, as with a contract clause.
04
Measuring quality
Common automatic scores such as ROUGE compare a summary with a reference summary. The 2020 study found that textual entailment measures, which ask whether the source supports each statement, correlate better with faithfulness than the standard metrics. For an office the practical tests are faithfulness (is each claim supported?), coverage (what is missing?) and time saved (including the time spent checking).
05
Where ANVE stands
ANVE’s minutes maker ties every action item to the transcript line it came from, and its document question answering shows the source passage, so a reader can check an answer against the original. Full document drafting with design templates is not on ANVE yet.
06
Questions people ask
Can AI summarise a 100-page PDF?
Often, but check the middle: research shows language models use the start and end of long inputs better than the middle. Summarising section by section and keeping page references reduces the risk. Our question page covers it.
Is an AI summary safe for legal documents?
Treat it as a map, not the text. Read the clause that matters in the source before acting. Our question page on mistakes in legal documents explains why.
What is hallucination in a summary?
Content the source does not support. It is either a distortion of something the source says, called intrinsic, or information the source does not contain, called extrinsic.
Should I use extractive or abstractive summaries?
Use extracted quotes for numbers, dates and names, and abstractive text for readability, and show both so the reader can check one against the other.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
