survey-seed-harvest_skill

Overview

This skill bootstraps taxonomy seeds by identifying survey and review papers in a deduplicated retrieval set and extracting topic/subtopic terminology into outline/taxonomy.yml. It produces a concise, actionable seed taxonomy (top-level chapters, children, short descriptions, and example titles) intended as input for a downstream taxonomy-builder. Treat outputs as starting points that must be rewritten and scope-aligned.

How this skill works

The skill scans papers/papers_dedup.jsonl for likely surveys using keywords like survey, review, systematic, and meta-analysis. It extracts frequent candidate terms, groups them into ~4–10 top-level nodes with 2–6 child leaves, and generates short descriptions indicating what belongs here and what does not. The result is written as YAML seeds for taxonomy-builder to refine and align with project scope.

When to use it

• After retrieval/dedup when you want a fast initial topic structure from existing surveys
• When bootstrapping a domain taxonomy to accelerate human curation
• If you have multiple review papers and need candidate chapter/section labels
• Before running taxonomy-builder to provide concrete starting seeds
• When you prefer to reuse field-established frames from surveys (and will rework them)

Best practices

• Treat output strictly as seeds; run taxonomy-builder to rewrite and align scope
• Prefer concrete terms (methods, datasets, tasks) over generic placeholders
• Verify that every node has a clear description and at least two levels
• If no surveys are found, widen retrieval keywords or add known survey papers
• Use conservative term thresholds to avoid noisy, low-frequency terms

Example use cases

• Extract chapter-like seeds from 10 retrieved survey papers to speed ontology design
• Generate leaves of methods, benchmarks, and evaluation criteria from reviews
• Produce representative titles under each seed node for human reviewers
• Quickly populate outline/taxonomy.yml before a manual taxonomy-building sprint
• Recover domain terminology after deduped retrieval for downstream tagging