Structural Alignment and Semantic Divergence in Culturally Configured AI Evaluation of Scientific Titles Using BERT and QAP Network Analysis

AI-assisted scholarly communication must work reliably across cultures, and this study tests whether high-context and low-context AI evaluators construct different meanings from identical scientific titles. We examine how scientific titles are interpreted by AI agents simulating high-context (HC) and low-context (LC) communication styles. Drawing on Hall's (1976) framework, participants rewrote 178 peer-reviewed journal titles from engineering, computer science, and management using metaphorical, idiomatic, and culturally implicit strategies. These original and enriched titles were evaluated by two culturally configured AI mentors to investigate how cultural framing shapes interpretation. We employed BERT-based semantic modeling and Quadratic Assignment Procedure (QAP) network analysis to assess both structural convergence and semantic divergence in AI-generated evaluations. While HC and LC agents showed strong structural alignment (QAP r=.65; R²=.42), semantic similarity was negligible (r=-0.011). HC interpretations emphasized metaphorical clustering and holistic coherence, while LC agents favored lexical clarity and conceptual modularity. The findings highlight how culturally informed evaluators construct divergent meanings from identical texts and extend cultural discourse theory into AI-mediated evaluation.