Integrating syntactic covariates into topic modeling : using universal dependencies for improved high-level summaries of customer reviews

Kurzfassung/Abstract

Marketing text analysis often relies on narrow, specialized datasets: too limited for generic large-scale NLP, yet too large for manual review. This dissertation introduces labeled UD-LDA, a topic model that conditions topic propagation on Universal Dependencies, allowing topics to flow along syntactic relations rather than relying solely on word order or co-occurrence. Applied to three customer review datasets, it outperforms standard benchmarks, yielding better model fit and more distinct yet coherent topics. Modeling propagation as a function of dependency type reveals that modifiers and function words promote topic consistency, while relations linking distinct syntactic units suppress it, showing that grammar systematically structures latent thematic content.