Thematic Modeling of Technical Support Requests

Interest in the tasks of generalization, highlighting the meaning from a text or a collection of texts is steadily growing. Topic modeling or forming set of keywords characterizing a text is increasingly being used in scientific research and business tasks. This paper demonstrates an approach in solving practical business problem - highlighting the key topics for technical support service requests in the corporation in 2023-2024. There are two approaches used, first one – the LDA which is considered to be the classical in terms of machine learning and the second one which is using transformer models, namely Bertopic to address the business problem. For classical mechanisms (LDM models), we consider the ways to increase the quality of the model by preprocessing the data and tuning model hyper parameters. For modern (Bertopic\HDBSCAN) mechanisms, both unsupervised and reinforcement learning approaches are used. As an addition to the business problem, also the possibility of tracking the appearance of new bulk topics among the requests is considered, which is of particular value in the context of the task of clustering requests to the technical support service. Keywords: LDA (Latent Dirichlet Allocation), Bert, Bertopic, topic modeling, technical support requests, key topics