Orthogonal rotations in latent semantic analysis: An empirical study

•Orthogonal rotations influence factor interpretation in Latent Semantic Analysis.•Varimax and Equamax produce factors with similar interpretation.•Quartimax tends to summarize the content of text corpora in a single extracted factor.

The Latent Semantic Analysis (LSA) literature has recently started to address the issue of interpretability of the extracted dimensions. On the software implementation front, recent versions of SAS Text Miner ® started incorporating Varimax rotations. Considering open source software such as R, when it comes to rotation procedures the user has many more options. However, there is a little work in providing guidance for selecting an appropriate rotation procedure. In this paper we further previous research on LSA rotations by introducing two well-known orthogonal rotations, namely Quartimax and Equamax, and comparing them to Varimax. We present a study that empirically tests the influence of the chosen orthogonal rotations on the extraction and interpretation of LSA factors. Our results indicate that, in most cases, Varimax and Equamax produce factors with similar interpretation, while Quartimax tends to produce a single factor. We conclude with recommendations on how these rotation procedures should be used and suggestions for future research. We note that orthogonal rotations can be used to improve the interpretability of other SVD-based models, such as COALS.

Graphical abstractFigure optionsDownload full-size imageDownload as PowerPoint slide