What problem does it solve? Systematic literature searches often return thousands of records contaminated by wildcard overflow, homonyms, and weakly related citations. Manually screening each record is slow and inconsistent, and one-pass keyword filtering produces both false positives and false negatives. ## Core Features & Use Cases - Three-Layer Noise Model: Distinguishes word-form noise (wildcard overflow), topic noise (homonyms), and boundary noise (weak associations) with dedicated judgment paths for each. - Multi-Dimensional Relevance Framework: Evaluates word meaning, research topic, concept role, discipline, and knowledge contribution to assign verdicts like RELEVANT, NOISE, or graded PENDING levels. - Iterative Multi-Round Cleaning: Processes literature in batches of 20-50 records, writes verdicts back to a mapping table, refines noise rules between rounds, and stops when noise density falls below 1%. - Use Case: A researcher running a bibliometric study on memetics retrieves 5,000 records via a meme* wildcard query. This skill screens them batch by batch, filtering out unrelated hits (e.g., internet memes in engineering contexts) while preserving genuinely relevant cultural-evolution studies. ## Quick Start Ask the AI to start round 1 semantic screening of the full literature dataset using the prepared mapping table and noise expectations.
