Two-step semantic clustering of embeddings as an alternative to LDA for infometric analysis of industry news.

The subject of the research is the development and validation of an alternative approach to thematic modeling of texts aimed at overcoming the limitations of classical Latent Dirichlet Allocation (LDA). The object of the study is short Russian-language news texts about nuclear energy, presented in the form of the "AtomicNews" corpus. The authors thoroughly examine various aspects of the topic, such as the impact of sparsity on the quality of thematic modeling, issues of theme interpretability, and the limitations of a priori fixing the number of topics. Special attention is paid to the geometric interpretation of text semantics, in particular, the transformation of lexical units into the space of pre-trained embeddings and subsequent clustering aimed at forming document thematic profiles. The research focuses on the comparative analysis of the new method and LDA using coherence, perplexity, and thematic diversity metrics. The proposed approach aims to create an interpretable, computationally lightweight, and noise-resistant model suitable for online monitoring of news flows. The research methodology is based on a two-stage semantic smoothing process—embedding representation of lemmas using Sentence-BERT and agglomerative cosine clustering, followed by the application of K-means to the thematic profiles of documents. The scientific novelty of the study lies in the development and empirical justification of a thematic modeling scheme that replaces probabilistic word generation with geometric smoothing of embeddings. The proposed approach departs from the assumptions of the "bag of words" and a fixed number of topics, forming thematic coordinates of documents through density clusters in semantic space. This enhances theme interpretability, reduces sensitivity to text sparsity, and avoids the collapse of topic distribution in short messages. Experiments on the "AtomicNews" corpus demonstrated a statistically significant improvement compared to classical LDA: a 5% reduction in perplexity, a 0.15-point increase in topic coherence, and an increase in thematic diversity. The method also demonstrated computational efficiency—the entire procedure takes seconds on a CPU, making it suitable for application in resource-constrained environments. Thus, the transition from probabilistic decomposition to geometric analysis of embeddings represents a promising direction in thematic modeling of industry texts.