Source-Linked Retrieval for Context-Dependent Short Clauses in Digitized Chemical-Safety Standards

Abstract

Digitizing standards does not by itself make their provisions reliably retrievable. Many requirements are expressed as short clauses whose regulated object, scope, list context, table value, or referenced condition is supplied elsewhere in the document. This study examines that indexing problem in a Neo4j deployment containing 9206 processed safety-related standards, 991,453 Sections, and 70,016 Tables. A repaired provenance layer maps 248,454 historical QAPair source records to unique Sections and exposes 1,207,059 question variants as attributable retrieval entries; 32,714 unresolved source records remain quarantined. We evaluate a source-linked design that searches Sections and QAPair fields in parallel and applies citation or dependency traversal only after query-time gating. On a frozen 120-query confirmation set, Section BM25 achieved Recall@10 of 0.792, deterministic context-enriched Section BM25 achieved 0.875, and QAPair question-plus-answer BM25 achieved 0.933. Question-only and answer-only Recall@10 were 0.833 and 0.900, respectively, showing that the gain cannot be attributed to question expansion alone. BGE-M3 Section retrieval, QAPair-question retrieval, and their dense fusion reached 0.800, 0.833, and 0.867, while BM25 top-100 retrieval followed by BGE reranking reached 0.908. On 120 ordinary queries not preselected for relations, the graph gate activated for 5 queries, improved none, and lowered one source rank; on 18 relation-dependent queries, the same gate increased complete two-clause recovery from 5/18 to 11/18. Two independent chemical-safety experts additionally reviewed selected context-restoration, query, graph, and end-to-end cases. These results support source-linked multi-entry retrieval as a bounded mechanism for this corpus while showing that deterministic context, answer fields, and selective graph activation must be evaluated separately.