C89 - Data Collection and Data Estimation Methodology; Computer Programs: OtherReturn

In this paper, we focus on the categorization of tickets in service desk systems. We employ modern neural network-based artificial intelligence methods to improve the performance of current systems and address typical problems in the domain. Special attention is paid to balancing the ticket categories, selecting a suitable representation of text data, and choosing a classification model. Based on experiments with two real-world datasets, we conclude that text preprocessing, balancing the ticket categories, and using the representations of texts based on fine-tuned transformers are crucial for building successful classifiers in this domain. Although we could not directly compare our work to other research the results demonstrate superior performance to similar works.

The paper presents a procedure that helps in revealing topics hidden in large collections of textual documents (such as customer reviews) related to a certain group of products or services. Together with identification of the groups containing the topics the lists of important expressions is presented which helps in understanding what characterizes these aspects most typically from the semantic point of view. The procedure includes determining an appropriate number of groups representing the prevailing topics, partitioning the documents into a desired number of groups using clustering, extracting significant typical features of documents from each group with application of feature selection methods, and evaluating the outcomes with the assistance of a human expert. The results show that the presented approach, consisting mostly of automated steps, is able to separate and characterize the aspects of a certain product as discussed by the customers and be later useful, e.g., for handling customer complaints, designing promotional campaigns, or improving the products.