Taxonomy Aligned Topic Modeling in Multimodal Systems Using Tactm++ : A Transformer-Based Topic Modeling Approach

Exponential growth of multimodal systems has seen a burgeoning of these strategies in the areas of fusion with many being defined in unstructured and domain-specific terms such that automated classification and semantic meaning extraction is extremely difficult. This research article presents TACTM++ (Taxonomy-Aligned Contextual Topic Modeling ++), a novel, modular architecture to attempt the explicit discovery of latent semantics of descriptions of fusion strategies and their alignment with canonical fusion taxonomies, such as early, late, hybrid, attention-based, graph-based, and so on. TACTM++ uses domain-adaptive transformer embeddings, self-supervised semantic clustering (UMAP + HDBSCAN), attention-based taxonomy alignment, and graph-based topic refinement to provide sharable topic models with high interpretability and good coherence and accurate category alignment. Large-scale inference experiments on synthetic multimodal corpora confirm that TACTM++ is superior to state-of-the-art methods in application to topics (LDA, BERTopic, Top2Vec, and Graph-Enhanced Topic Models) in both topic coherence (C v = 0.68), alignment precision (87.2 percent) and cluster quality. This architecture provides a flexible and scalable system of intelligent study of technical literature and the possibility of extracting insights on great scale (in heterogeneous modalities and fields of study).