Matching Meaning at Scale: Evaluating Semantic Search for 18th-Century Intellectual History through the Case of Locke

Researchers evaluate semantic search as a tool for analyzing 18th-century intellectual history, specifically tracking how John Locke's ideas circulated through paraphrases and implicit references. While semantic search substantially outperforms traditional lexical methods at capturing meaning-level correspondences, linguistic analysis reveals that retrieval remains constrained by surface-level vocabulary overlap, suggesting both promise and limitations for historical corpus analysis.

This research addresses a fundamental challenge in digital humanities: detecting how ideas evolve and spread through texts when authors paraphrase rather than quote directly. Traditional lexical text reuse detection captures only verbatim quotations, missing the majority of intellectual transmission that occurs through paraphrasing and implicit engagement. The study uses semantic search—powered by embedding models—to detect meaning-level correspondences in 18th-century texts discussing Locke, finding substantially more implicit receptions than keyword-matching approaches.

The work contributes to broader efforts in computational humanities and information retrieval by empirically demonstrating semantic search's effectiveness while honestly confronting its limitations. Expert annotation grounded in semantic taxonomy provides rigorous evaluation standards, moving beyond purely automated assessment. The discovery of "lexical gatekeeping"—where even semantic retrieval remains partially constrained by vocabulary overlap—reveals that current embedding models inherit linguistic biases from training data.

For the AI research community, these findings underscore a key limitation in semantic understanding: models struggle to fully decouple meaning from surface form, particularly in specialized historical contexts. This has implications for any domain requiring deep semantic matching across diverse vocabularies. The research demonstrates that while neural semantic search represents genuine progress over lexical baselines, practitioners should not assume embedding models achieve true semantic understanding independent of linguistic features.

Looking forward, advances in domain-specific embeddings, multimodal representations, and hybrid retrieval systems may address lexical gatekeeping. The open dataset enables broader community validation and development of improved semantic search methods for historical corpus analysis.

• →Semantic search detects substantially more implicit intellectual transmission than lexical text reuse methods in historical corpora.
• →Current embedding models exhibit 'lexical gatekeeping,' remaining partially constrained by surface vocabulary despite semantic capabilities.
• →Expert annotation with semantic taxonomy provides rigorous evaluation beyond automated metrics for historical text analysis.
• →Domain-specific semantic search requires addressing vocabulary biases inherited from training data to achieve true meaning-level retrieval.
• →Hybrid approaches combining semantic and lexical methods may prove more effective than semantic-only solutions for specialized historical analysis.