Topic modeling of chat data: A case study in the banking domain

Abstract

Text analytics and natural language processing become increasingly important in many business companies and the financial services world is no exception. The banking sector is struggling to find new ways of communicating with customers in order to answer to their needs in various innovative ways. The aim to ease customer communication through modern technologies has a central place in the digitalization roadmap of many banks and here comes the important role of text analytics. The goal of this paper is to derive new and useful insights on customer behavior, study the usefulness of a well-known topic modeling technique and propose a methodology for its application in the banking domain. The current empirical study is carried out on real chat data generated between operators in a call center and clients of a large bank. We capture the main topics of client interest, analyze their characteristics and trace their development over time. Several challenges are faced during the empirical study mainly stemming from the domain and noisiness of data - it is predominantly in Bulgarian but there are chats in other languages too; transliteration is extensively used; there are some specific structural characteristics inherent to chat data. Nevertheless, results are promising - the extracted topics are interpretable and meaningful which is confirmed by the application of different techniques for evaluation of results. As a major contribution of the current empirical study, we consider the valuable insights into customer behavior obtained from the analysis of real-world chat data in the banking domain. Furthermore, results can be used not only for tracking trends in customer behavior, but they can serve as guidance for the effective development of chatbot systems. Considering the challenges imposed by chat data in the banking domain, another contribution of the current paper is the development of a methodology for topic extraction from chats in this domain. The current research also suggests what text processing techniques should be applied on customer support chats in order to obtain more meaningful results. The last is applicable to data not only in the banking domain, but in any other text data domain too. Finally, results from the experiment are also evaluated through the lens of the COVID-19 crisis and the analysis of its impact on customer behavior.