Traditional data mining techniques, while effective for structured numerical data, struggle with the complexity, ambiguity, and dynamic nature of natural language. To address these challenges, topic models have emerged as a powerful tool for uncovering latent topic structures in text, enabling applications such as trend analysis, recommendation systems, and decision support. These models are particularly valuable in social network content analysis, where they help reveal user behavior, information dissemination patterns, and social dynamics by identifying implicit topic distributions. The linguistic diversity of text data lowers the quality of the topic models. To overcome this shortcoming, Transformer-based text encoding is applied. However, benchmarks for text encoding are not set, and the models have not been properly compared in terms of dimensionality/noise reduction. We conduct comparative analysis of five neural-network encoding models based on the Transformer architecture, integrating them with clustering methods for efficient topic detection. We employ a dataset of COVID-related discussions from Twitter collected between November 24, 2021, and January 23, 2022. The methodology pipeline includes text encoding, dimensionality reduction, and clustering, with detailed steps and results discussed in the respective sections. The results show that the encoding models differ substantially in how they help in topic detection, with the general text embedding (GTE) model providing for the best quality. Our findings highlight the models’ differing efficiency in topic detection, offering insights for future research and practical applications in text analysis and knowledge discovery.