BERT-guided nested Dirichlet forests sharpen hierarchical topic modeling

Researchers have introduced a hierarchical topic model that replaces the flat structure of classical Latent Dirichlet Allocation with a nested Dirichlet distribution as the document-topic prior, allowing topics to be organized in layers rather than in a single undifferentiated plane. The model, called DF_sLNDA, is described in a paper published on 25 October 2026 in the International Journal of Data Science and Analytics by Fares Alkhawaja, Manar Amayri and Nizar Bouguila. The authors argue that LDA, despite its enduring popularity, falls short in several respects because of its flat structure and its lack of semantic correlations among words, and their framework is designed to address both weaknesses at once.

The technical core of the approach is twofold. First, swapping the Dirichlet prior for the nested Dirichlet distribution permits hierarchical topics, so that broad themes can contain narrower subthemes in a structured nesting. Second, the model places a Dirichlet forest as a prior over these nested distributions, and the nestings are constructed from BERT-based semantic relationships. This means the tree-like structure of the forest encodes domain knowledge about which words belong together semantically, serving as a prior over the topic-word distributions rather than leaving the vocabulary relationships entirely unmodeled.

To make such a rich model tractable, the authors derive a complete mean-field variational inference procedure with closed-form updates for efficient parameter estimation. A further practical barrier to knowledge-guided topic models has been that users must typically specify domain constraints by hand. The paper addresses this with an automated pipeline built on CoSBERT, which discovers must-link and cannot-link constraints from the data itself, indicating which terms should be encouraged to share a topic and which should be kept apart.

On top of this, the researchers propose a TF-IDF-based adaptive weighting approach applied to the entire vocabulary, which controls the importance of individual terms within the discovered constraints. According to the abstract, this weighting step yields an additional 20 to 30 percent boost across all metrics.

Experiments on three benchmark datasets compare DF_sLNDA against standard LDA and report improvements of 75.6 percent in perplexity, 36 percent in coherence, and 27 percent in F1-score. The authors state that standard deviations were reported across five independent runs with different random seeds, a measure of how consistently the gains held across repeated estimation. Beyond the headline comparisons, the paper includes ablation studies that isolate the individual contributions of the TF-IDF weighting component and the BERT-based constraint component, showing what each element adds independently to the overall performance. To support reproducibility, the authors have released the full implementation of their framework, including the source code and the datasets used in the experiments, through a public anonymous repository.

Subject of Research: Hierarchical topic modeling using nested Dirichlet distributions with BERT-derived semantic constraints

Article Title: Adaptive BERT-based guided hierarchical topic modeling via nested Dirichlet forests

Article References: Alkhawaja, F., Amayri, M., & Bouguila, N. (2026). Adaptive BERT-based guided hierarchical topic modeling via nested Dirichlet forests. International Journal of Data Science and Analytics, 22(1), Article 355. https://doi.org/10.1007/s41060-026-01327-y

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01327-y

Keywords: topic modeling, LDA, nested Dirichlet distribution, BERT, Dirichlet forest, variational inference, CoSBERT, TF-IDF, natural language processing, machine learning, Adaptive, BERT-based