S04-映射表构建

Iterate through the raw export file and create one row per record with an incrementing seq number, the platform's unique source ID (UT/DOI/PMID), title, year, and document type. Then fill concept group and hit word fields from search-term analysis, leaving verdict fields empty for later rounds.

Decompose the search query into independent concept groups (e.g., G1, G2, G3), then test each record against every group. Record which groups matched in the concept_group field and the exact matched word forms in the hit_words field.

Core fields are seq (immutable ID), source_id, title, publication year, document type, concept_group, hit_words, verdict, noise_type, reason, and round. Extended fields like verdict_history and per-round noise types preserve revision traces across cleaning rounds.

No. Later rounds load the existing table and only update verdict, noise_type, reason, and round fields in place. When revising an earlier judgment, note the history in the reason field, and never modify the seq identifier.

Check that row count equals the raw record count, seq values run continuously from 1 to N without gaps or duplicates, source_id values are unique, and the file is saved as UTF-8 with BOM encoding.