A graph-based approach to topic clustering of tourist attraction reviews

Abstract

A large volume of user reviews on tourist attractions can prohibit travel businesses from acquiring overall consumers’ expectations and consumers themselves from seeing the big picture and making thoughtful decisions on trip planning. Summarization of the reviews allows both parties to catch the main themes and underlying tones of the attractions. In this paper, we address the task of topic clustering, by applying a graph-based approach to group the reviews into clusters. To interpret the resulting review clusters, WordNet and Inverse Document Frequency (IDF) are utilized to extract keywords from each cluster which represents the topic. We evaluate the graph-based clustering approach against gold standard data annotated by human and the results are compared against Latent Dirichlet Allocation (LDA), a widely used algorithm for topic discovery. The approach is shown to be competitive to LDA in terms of clustering user reviews on tourist attractions. The graph-based approach, unlike LDA which requires the number of clusters as an input, can dynamically clusters the reviews into groups, revealing the number of clusters.