Unveiling topic dependencies through a multilevel topic model: a hierarchical approach to enhanced interpretability

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Data Mining

Background:

• Topic modeling identifies key themes in unstructured text data.
• Latent Dirichlet Allocation (LDA) is a common topic model but fails to capture inter-topic dependencies.
• Understanding topic relationships is crucial for deeper text analysis.

Purpose of the Study:

• To propose a novel multilevel topic model (MTM) that captures hidden topic dependencies.
• To allow word-based topic proportions to vary across a multilevel latent structure.
• To improve the interpretability of topic modeling results.

Main Methods:

• Developed a multilevel topic model (MTM) with a multilevel latent structure.
• Employed a modified Expectation-Maximization (EM) algorithm with an upward-downward approach for parameter estimation.
• Conducted empirical studies on corpora to validate the MTM.

Main Results:

• The MTM successfully unearths hidden topic dependencies within a corpus.
• Word-based topic proportions were shown to vary across the higher-level latent structure.
• The multilevel hierarchy of the MTM was interpreted through empirical analyses.

Conclusions:

• The proposed multilevel topic model (MTM) offers superior systematic interpretability compared to Latent Dirichlet Allocation.
• MTM effectively reveals the inherent relationships and hierarchical structure between topics in text data.
• This model provides a more nuanced understanding of thematic structures in unstructured text.