The digitization of vast historical archives, such as the Eighteenth Century Collections Online (ECCO), offers unprecedented opportunities to trace the dissemination of ideas. However, traditional computational approaches to reception history rely heavily on lexical text reuse detection, which faces the challenges of implicit reference, significant OCR errors, diachronic semantic change, and the conceptual complexity of intellectual language. This challenge is particularly acute when tracing John Locke, whose complex arguments often appear in implicit paraphrases that evade lexical retrieval. To address this, we investigate semantic search as a complementary tool to close reading, evaluating its utility for tracking the reception of complex ideas. This study presents the first large-scale validation of semantic retrieval on ECCO, pioneering a paradigm beyond lexical reuse.
Our methodology utilizes the entire ECCO corpus (32 million pages), enriched with ESTC metadata. We segmented the corpus into 100-token chunks and encoded them using a Sentence-BERT model, constructing a FAISS index for efficient dense vector retrieval at scale. We implemented a multi-stage evaluation, beginning with the selection of 20 candidate quotes from a pool of 1,000 frequent reuses. For each candidate, we retrieved the top 200,000 semantically similar chunks, filtering out lexical reuses to isolate novel matches. Following a stratified manual annotation of 50 hits per quote, we selected two high-performing quotes for deeper analysis, considering both high retrieval quality and conceptual richness. We annotated a denser sample of the top 200 hits each, yielding two curated datasets of approximately 400 validated instances in total.
Our initial analysis reveals that while the relationship between similarity predictions and semantic categories is noisy, the signal-to-noise ratio remains favorable. Crucially, in 15 of 20 sampled quotes, the top-ranked hits were predominantly valid matches. In the further macro-level analysis of 400 curated instances, we found that the volume of works containing semantic reuse significantly exceeds that of lexical reuse. The distribution of works varies notably by genre and time; for instance, one quote shows a distinct peak in the 1740s within the science genre, linked to the rise of Newtonianism.
This study offers both a methodological proof-of-concept and a rigorous empirical resource. We demonstrate that a workflow combining semantic search with efficient annotation can surpass manual retrieval to discover valuable non-lexical hits. Future work will leverage these datasets as a gold standard to develop LLM-assisted annotation pipelines, allowing us to scale from single quotes to broad intellectual traditions.