Cluster Analysis of the Newcastle Electronic Corpus of Tyneside English: In A Comparison of Methods

This article examines the feasibility of an empirical approach to sociolinguistic analysis of the Newcastle Electronic Corpus of Tyneside English using exploratory multivariate methods. It addresses a known problem with one class of such methods, hierarchical cluster analysis—that different clustering algorithms can yield different analyses of the same data set, and that there is no obvious way of selecting the best one. The proposed solution is to analyze the data using hierarchical methods in conjunction with one or more fundamentally different types of clustering method, and then to select the analysis on which the hierarchical and the other method(s) agree most closely. A dimensionality reduction method, the self-organizing map (SOM), is used to exemplify this approach. The result is a close though not perfect match between the SOM and complete-link hierarchical analyses, but there is an important reservation—the SOM results vary with changes in user-defined training parameters, and are consequently also open to the criticism of inconsistency. The SOM cannot therefore be an objective arbiter for hierarchical clustering, but the analysis on which they agree gives a better basis for understanding the structure of the data than either method can provide on its own