A topic change point model for phrase-based topic inference

Abstract

Obtaining customer insights from large and unstructured text corpora (e.g. customer reviews, blogs, tweets, written exchange in user forums or customer service emails) has attracted significant interest in marketing research. Topic models are among the most popular descriptive devices to analyze such data. We propose a new type of topic model built on observed word sequences. The proposed topic change point model assumes that text consists of sequences of words belonging to the same topic, possibly interspersed by ubiquitous terms such as stop words, and that topics change from run to run so that the end of each topic run marks a topic change point. Our model yields strings of words assigned to the same topic for phrase-based topic inference. In an extension of our model, we use parts-of-speech tags as prior information to topic change points which allows for observed syntactical structure of text to enter topic inference. We apply our model to two data sets and compare it to alternative modeling approaches, including state-of-the-art, BERT based topic models. We investigate the capability of models to predict consumer ratings which addresses their power in summarizing words so that the origin of ratings can be assessed by marketers. Implications and directions for future research are discussed.