Exploring performance of clustering methods on document sentiment analysis

Abstract

Clustering is a powerful unsupervised tool for sentiment analysis from text. However, the clustering results may be affected by any step of the clustering process, such as data pre-processing strategy, term weighting method in Vector Space Model and clustering algorithm. This paper presents the results of an experimental study of some common clustering techniques with respect to the task of sentiment analysis. Different from previous studies, in particular, we investigate the combination effects of these factors with a series of comprehensive experimental studies. The experimental results indicate that, first, the K-means-type clustering algorithms show clear advantages on balanced review datasets, while performing rather poorly on unbalanced datasets by considering clustering accuracy. Second, the comparatively newly designed weighting models are better than the traditional weighting models for sentiment clustering on both balanced and unbalanced datasets. Furthermore, adjective and adverb words extraction strategy can offer obvious improvements on clustering performance, while strategies of adopting stemming and stopword removal will bring negative influences on sentiment clustering. The experimental results would be valuable for both the study and usage of clustering methods in online review sentiment analysis.