Presentation Information
[3Yin-A-12]Comparative Evaluation of Evidence-Linked Structured Summaries for Administrative Long-Term Care Q&A
〇Seiichirou Yokoo1, Yoriko Yokoo1 (1. Terra Incognita LLC)
Keywords:
Evidence-supported summarization,Long-term care insurance Q&A (Japan),Manual and automatic evaluation,Information extraction,Structured summarization
Administrative Q&A documents are often presented in tabular form and contain long, exception-rich descriptions, making evidence-preserving summarization difficult. We converted the Japanese Ministry of Health, Labour and Welfare Long-Term Care Service Q&A (XLS) into JSONL and compared (i) a plain summarization baseline and (ii) and evidence-linked structured summarization that decomposes the output into rule / exception / procedure statements (S1, S2, ...) and attaches and evidence ID to each statement. From 2,028 deduplicated items, we generated 100 stratified samples by document length. The baseline failed format / structure requirements in three cases, while the proposed method failed in one case; the remaining 96 pairs were evaluated by two human evaluators. Both evaluators preferred the proposed method more often (50 / 42) than the baseline (9 / 18), with 37 / 36 ties. These results suggest that evidence-linked structured summaries improve usability while preserving auditability in policy Q&A.