When Does Evidence Expansion Help? A Cross-Domain Study of Retrieval Paths and Lightweight Routing for Regulatory Evidence
International Journal of Computer Science and Information Technology · 30.10.2026 · 8 (7) · 63-75
DOI: 10.62051/ijcsit.v8n7.07 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Computer Science and Information Technology, online erschienen am 30.10.2026. Verfasst von Kaifeng Sun. Die Arbeit nennt 23 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
Regulatory question answering often fails because the decisive evidence is not contained in the retrieved leaf passage but in its heading, parent clause, table, or cited provision. We study whether common evidence-expansion stages improve retrieval before generation under a risk-sensitive criterion that requires complete supporting evidence and rejects materially misleading additions. In a prespecified pilot of 120 questions drawn from Chinese chemical-safety standards and English pharmaceutical-regulatory documents, we compare six retrieval paths built on a shared BM25 first stage, including reranking, structural-context expansion, relation expansion, and their combinations. Human annotators labeled 10,385 evidence occurrences for completeness and harmfulness. Structural context provided the clearest benefit: adding heading, parent, and table context improved combined success by 15.83 percentage points over BM25, whereas reranking and one-hop relation expansion yielded limited independent rescue. The most complete path did not outperform the cheaper context-only path because additional evidence often introduced harmful material. Lightweight learned routers outperformed a prespecified heuristic in grouped out-of-fold evaluation, but calibration-based abstention failed to achieve usable risk and coverage. These results support selective use of structural context and separate evaluation of completeness and harm, while indicating that reliable evidence-sufficiency abstention requires substantially more calibration data.
Zitieren
APA
Sun, K. (2026). When Does Evidence Expansion Help? A Cross-Domain Study of Retrieval Paths and Lightweight Routing for Regulatory Evidence. International Journal of Computer Science and Information Technology, 8(7), 63-75. https://doi.org/10.62051/ijcsit.v8n7.07
Vancouver
Sun K. When Does Evidence Expansion Help? A Cross-Domain Study of Retrieval Paths and Lightweight Routing for Regulatory Evidence. International Journal of Computer Science and Information Technology. 2026;8(7):63-75. doi:10.62051/ijcsit.v8n7.07
BibTeX
@article{Sun2026,
author = {Sun, Kaifeng},
title = {When Does Evidence Expansion Help? A Cross-Domain Study of Retrieval Paths and Lightweight Routing for Regulatory Evidence},
journal = {International Journal of Computer Science and Information Technology},
year = {2026},
volume = {8},
number = {7},
pages = {63--75},
doi = {10.62051/ijcsit.v8n7.07}
}
EUVIMED