Thematic alignment across expert-guided and machine-derived topic models in Positive Energy District characterization

Positive Energy Districts (PEDs) are increasingly framed through both expert-driven policy narratives and data-driven analyses of emerging project discourse. However, the extent to which different knowledge extraction approaches produce conceptually meaningful representations of the PED concept remains insufficiently examined. This study examines thematic patterns in two representative English-language corpora using expert-guided and machine-derived topic modeling with a three-way analytical design. A Pseudo-document-Guided LDA (PG-LDA) model was applied to structured case descriptions from the PED-DB (66 documents after preprocessing), while unsupervised Latent Dirichlet Allocation (LDA) was applied to an unstructured corpus of 35 documents retained from 44 identified PED project sources (approximately 149,420 words). To provide a comparative reference point for interpreting differences attributable to modeling approach and corpus type, an additional unsupervised control LDA (K = 7) was trained on the same structured corpus as PG-LDA. Cross-corpus semantic alignment was evaluated using sentence-embedding cosine similarity (SBERT; all-MiniLM-L6-v2), which accommodates the near-zero lexical overlap between expert-defined bigram topics (PG-LDA) and corpus-derived unigram topics (LDA). Results indicate moderate semantic alignment across most thematic dimensions (SBERT cosine similarity range: 0.348–0.535), with the cross-model comparison (PG-LDA vs. control LDA; mean SBERT = 0.453) and the cross-corpus comparison (control LDA vs. unstructured LDA; mean SBERT = 0.462) showing broadly comparable alignment scores. Stakeholder integration showed consistently weak alignment across all model comparisons, while financial feasibility and mobility integration remained comparatively underrepresented in both corpora. These asymmetries suggest differences in how PED knowledge is produced, communicated, and institutionalized across expert-curated and project-communication contexts. The findings suggest that expert-guided modeling supports conceptual consolidation of policy-relevant themes, whereas machine-driven modeling reveals emergent narratives and contextual experimentation dynamics. By integrating deductive and inductive analytical perspectives, the study contributes to a more reflexive understanding of PED discourse and supports the development of more balanced knowledge frameworks with broader qualitative and institutional evidence.