Large language models for scientometric mapping of scientific controversy: A validated hybrid AI–Human framework

Abstract This study advanced large language model (LLM)-enabled scientometric methods for analysing contested scholarly communication through a hybrid AI–human workflow for large-scale mapping of stance and themes in scientific corpora. We treated the literature as a large textual record and examined how competing knowledge claims were distributed across venues and over time. We used the Lyme disease controversy as a testbed, applied multiple LLMs to thousands of abstracts, and validated the outputs with domain experts to trace longitudinal shifts in authorial stance and thematic focus. We examined how alternative perspectives were distributed across venues and themes, how these distributions changed over time, and how venue stratification shaped the visibility of different epistemic positions. Results showed persistent asymmetries in publication placement and citation attention, with higher-impact venues disproportionately hosting one perspective relative to the other, while thematic mapping identified recurring fault lines around disputed claims. Validation showed substantial alignment between LLM outputs and expert ratings across classification tasks (Cohen’s kappa 0.58–0.71), broadly comparable to expert–expert agreement and sufficient for corpus-level scientometric analysis. The framework offers a reproducible model of human–LLM integration in literature-based scientometrics and enables venue- and time-resolved indicators for monitoring controversies and publication imbalances.