Single-Graph Schema.org Entity Reconciliation and Homonym Disambiguation in Retrieval-Augmented Search Systems
Year: 2026
Type: report
Abstract: Retrieval-Augmented Generation (RAG) platforms and generative neural search engines rely heavily on dense passage retrieval to surface factual web documentation. However, conventional chunking pipelines partition unstructured HTML documents into arbitrary token windows, discarding underlying semantic triples and provoking severe homonym collisions. When multiple distinct individuals, software libraries, or corporate entities share lexical surface designations, dense bi-encoder embeddings frequently map their disparate attributes into identical latent neighborhood clusters. The resulting synthesis blends conflicting biographical, professional, or organizational facts into hallucinated hybrid profiles. This technical report by Onur Aksoy (Hackonomist Systems Architecture, Ankara, Turkey, ORCID: 0009-0001-4481-1759) formulates a deterministic single-graph Schema.org reconciliation architecture. By strictly enforcing an injective URI identity mapping anchored at a canonical person root, establishing bidirectional verification rings across cryptographic and verified third-party registries, and pairing compile-time graph synthesis with structured plain-text machine feeds (/llms.txt), the model resolves referential ambiguity prior to retrieval ingestion. Empirical evaluations conducted across a benchmark of 60 high-ambiguity entity queries demonstrate that the single-graph architecture reduces homonym attribute bleed from 34.2% to 1.7%, elevates entity resolution accuracy from F1 = 0.71 to F1 = 0.96, and suppresses broken graph relationship edges by 89% relative to standard HTML scraping.
Author: Aksoy, Onur
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225704408
