본문 바로가기
  • Home

The Effect of Personal Ontology-Based Knowledge Graph Reranking on PKMS Search Quality

  • Journal of Korean Society of Archives and Records Management
  • Abbr : JRMASK
  • 2026, 26(2), pp.47~70
  • DOI : 10.14404/JKSARM.2026.26.2.047
  • Publisher : Korean Society of Archives and Records Management
  • Research Area : Interdisciplinary Studies > Library and Information Science > Archival Studies / Conservation
  • Received : April 8, 2026
  • Accepted : April 26, 2026
  • Published : May 31, 2026

Jung Hyeji 1

1컨텍스트에이 (ContextA) 대표

Accredited

ABSTRACT

This study investigates the efficacy of ontology-based search reranking that incorporates the producer’s cognitive dispositions and value systems within a Personal Knowledge Management System (PKMS). Grounded in the archival principle of provenance, the study hypothesizes that integrating a producer’s cognitive patterns and value systems can enhance the relevance of search results in personal records. To test this, POR (ver. 1.0–2.1.0), a system that reranks Gemini File Search RAG results using seven ontology scoring factors—hub boost, identity lens matching, cross-domain intersection, thinking pattern boost, graph proximity, value alignment, and authorship—was developed. The system was assessed on 3,702 markdown-based PKM notes, constructing a knowledge graph with 5,717 nodes and 13,544 edges, including 1,810 auto-linked HUB_CONNECTION edges. All 10 hub notes ranked within the top 15 by PageRank, supporting the validity of the hub selection methodology. In a live A/B comparison with 10 queries across 5 types, ontology-based reranking changed an average of 4.4 out of 10 positions, with a mean Kendall’s tau of 0.621 (SD = 0.252). Hub-related queries showed the largest reranking (tau = 0.300), while general queries preserved existing rankings (tau = 0.786). Identity lens matching (57 activations) and cross-domain intersection (29 activations) were the most frequently activated factors. These results demonstrate that personal ontology modeling provides contextual relevance beyond what embedding-based semantic similarity alone can achieve, with effectiveness varying by query type.

Citation status

* References for papers published after 2024 are currently being built.