Development of a Dialog Classification System for Scoring Systems of Banks

This paper discusses the use of natural language processing techniques, in particular dialog classification systems in the banking sector for further integration with classical scoring cards. The review considers the existing methods of natural language processing, the main language models for text classification problems as well as the state-of-the-art models on open datasets. The aim of the work is to improve the quality of CRM model scorecards by blending natural language processing models and tabular scorecards. During the study, the advantages and disadvantages of both individual methods and models and the current best solutions for this task were identified. As a result of the study, a new model architecture for processing dialog data based on transformers, LSTMs and convolutional neural networks is proposed. During the testing of the models’ results the proposed solution showed the best result on a bank dataset and is slightly inferior in quality to one of the best models on an open dataset. The developed model allowed to draw some conclusions about customer behavior, which are presented in the final chapter. The volume of the paper is 49 pages. The number of tables in the work - 4, illustrations - 22, used sources – 47.