Discovery of customer satisfaction dimension from tweets using latent dirichlet allocation

Abstract

Customer satisfaction of a service or product is reflected in the customer's attitude towards the service or product once the customer perceived or used the service or product. Customer satisfaction has many dimensions, and previous studies have identified and used those many dimensions to measure the customer satisfaction of a certain service or product. In this study, we explore the use of text mining techniques of modeling topic as a way of discovering customer satisfaction dimensions, and we used tweets data from Twitter as our dataset. These are our contributions. We believed that this approach had not been used for discovering customer satisfaction dimensions. We used Latent Dirichlet Allocation for the topic modeling, and perplexity model and topic coherence are used for optimizing the number of topics. We would then choose representative words for each topic formed using Latent Dirichlet Allocation. The name of the topic is added by analyzing the logical relationships between the representative words and find the semantic meaning of the representative words in the context of the original tweets. These are done manually by the authors. Once the labels have been identified, we associate the labels to the original tweets. The identified topics in our study become customer satisfaction dimensions. The last task in our study is comparing the identified dimensions with dimensions in previous studies. In this task, we needed to broaden the definition of dimensions from previous studies. The result shows our identified dimensions are fitted with the dimensions from previous studies. The main differences are the naming of the dimensions and the granularity of the dimensions. Our identified dimensions have finer granularity than previous studies. Our study enables the decision maker to be informed with the current dimensions, which are mattered by the customer in using the organization's services or products. Once the tweets are labeled to the identified topic names, we can sort the topics having the most documents. This result shows important issues which are discussed by the customers.