A Human-in-the-Loop Embedding Study of Textual Drift Using the InterLink Research Platform
This study evaluates semantic stability between the King James Version 1611 and KJV 1769 editions using an aligned-unit computational workflow developed for the InterLink Research Platform. The analysis compares 66 shared biblical works across 31,102 structurally aligned textual units using embedding-based cosine distance, robust thresholding, Monte Carlo permutation baselines, and human-in-the-loop annotation.
The results show that the KJV 1611 and KJV 1769 editions are highly stable semantically at the aligned-unit level. Across the shared corpus, median drift was low at 0.0592, with a Central Semantic Stability Index of 0.9408. A small upper-tail population of review candidates and strong outliers was identified, representing 5.13% of aligned units.
Human annotation showed that this drift tail was dominated by orthographic and character-form modernization rather than broad doctrinal or semantic divergence. In practical terms, many of the highest drift cases reflected early modern textual forms such as spelling, typography, capitalization, punctuation, ampersand expansion, and related character-level modernization.
The study validates InterLink as a reproducible workflow for computational theological text comparison and highlights the importance of combining embedding-based analysis with human interpretation when working with historical religious texts.
Paper and Reproducibility Package
Read the paper: Github Direct Link
View Reproducibility package: Zenodo DOI
Leave a Reply