Cross-Domain Multilingual Clustering: A Generative Hybrid Model for Constructing and Enhancing Semantic Graphs from Heterogeneous Data

Existing hierarchical clustering approaches face limitations in correctly representing semantic graphs and heterogeneous multilingual data, thereby limiting their ability to extract insights from complex datasets. The challenge is heightened by the complexity of integrating hierarchical neural topic-seeding models with semantic graphs for multilingual clustering, involving the refined combination of heterogeneous data and structural knowledge. In this context, we propose “the Multilingual Hybrid LDA” (MultiHLDA), a novel 5-phase generative cross-domain approach that integrates prior domain knowledge into hierarchical topic modeling. It involves multilingual distributional term clustering over Fundamental Concepts (FC) composing upper semantic graphs for semi-automatically learning and enhancing a Bottom-Up Universal Upper Semantic Graph (BU³SG) from heterogeneous data. MultiHLDA includes model fine-tuning and FC integration as seed terms to cluster semantically related terms into concepts. We aim to create high-quality and non-overlapping clusters by aligning term clusters with established FC and utilizing noun phrase patterns. Our approach synergizes data with semantic nuances using techniques from both graph mining and machine learning. The empirical findings of this approach emphasize its effectiveness in constructing and enhancing the BU³SG, highlighting its capacity to improve hierarchical topic seeding discovery, clustering, and document/graph representation. This notable progress extends to achievements against trained on ontology, fisheries, and medicine datasets.