Beyond citation-based metrics: Measuring interdisciplinarity via SBERT semantic embeddings and its heterogeneous effects on citation impact

Background: Interdisciplinary research is a cornerstone of global science policy, yet decades of research have reached conflicting conclusions about its association with citation impact. This inconsistency stems primarily from traditional indicators, which rely on reference diversity rather than genuine semantic knowledge integration, and from small, discipline-specific samples that limit generalizability. Objective: This study introduces a novel semantic interdisciplinarity measure based on Sentence-BERT (SBERT) embeddings, which directly captures cross-disciplinary knowledge integration at the textual level, and tests its heterogeneous relationship with citation impact across the full spectrum of scientific disciplines. Methods: We analyzed 121,194 articles published 2015–2025 across all 19 root-level disciplines in OpenAlex. We validated the reliability of OpenAlex disciplinary classification using multi-dimensional semantic analyses, and compared our SBERT-based indicator with the Simpson Diversity Index and Rao–Stirling Index. We employed OLS and negative binomial regressions with discipline and year fixed effects (standard errors clustered at the discipline level), journal tier heterogeneity analysis, and domain-specific decomposition analyses. Results: The semantic interdisciplinarity indicator shows moderate convergent validity with conventional citation-based metrics (r = 0.333–0.347, p