Summarization

Extractive methods select source passages, while abstractive methods generate new wording that condenses or combines information. A summary may cover one document, several sources or material selected by a question. Its quality depends on source fidelity, salience, coherence and the requested compression, which can conflict. Generative models condition on available text, so truncation or chunking determines which evidence can influence the result. Reference-overlap metrics compare wording with sample summaries but do not fully measure factual consistency. Faithfulness asks whether claims are supported by the source; factuality can also consider external truth. These are distinct, since a source itself may contain an incorrect claim.

Specify audience, purpose, maximum length and information that must survive compression. Preserve source provenance and ensure decisive sections reach the model. Compare a simple extractive baseline with generation or hierarchical processing. Keep documents and summary variants grouped across data splits. Evaluate coverage against a content checklist, inspect every important claim and test numbers, actors and caveats. Use reference metrics as supporting evidence rather than the release criterion. The deliverable is a reviewed summarization workflow with a clear source boundary and an evaluation rubric that makes omissions and unsupported additions visible.

Long inputs can lose crucial evidence through truncation or imperfect chunk aggregation. Extracted sentences can mislead when separated from qualifications, while generated wording can change causality or completion status. Multiple sources may disagree, and a summary must not silently merge their claims. Wording overlap does not establish fidelity. Summarization differs from open-ended generation and extraction of fixed fields. Check source support, coverage and reader usefulness separately, including whether the requested compression leaves enough space for essential uncertainty.