Predicting Customer Satisfaction in Service Processes Using Multilingual Large Language Models

Abstract

The huge amount of data recorded during business process executions in today’s organizations creates the need to leverage this data. While most existing business process monitoring methods are capable of including structured context information, the incorporation of unstructured information, for example, text, has rarely been researched. Recent advances in natural language processing offer tools to generate contextualized sentence embeddings, which capture more information in human language than ever before. Especially in service processes, such as the incident management process, a variety of unstructured information is created throughout the execution, representing a relevant use case for incorporating (multilingual) text. To close this research gap, a method exploiting multilingual text for predicting the outcome of a service process is presented. Multilingual large language models are used to generate sentence embeddings for unstructured text created during process execution. After instantiating the method, an evaluation was performed using a real-life event log to predict customer satisfaction in the incident management process at a German multinational corporation. We show that text incorporation improves predictive performance.