Deriving semantic structure from category fluency: Clustering techniques and their pitfalls

Abstract

Assessing verbal output in category fluency tasks provides a sensitive indicator of cortical dysfunction. The most common metrics are the overall number of words produced and the number of errors. Two main observations have been made about the structure of the output, first that there is a temporal component to it with words being generated in spurts, and second that the clustering pattern may reflect a search for meanings such that the ‘clustering’ is attributable to the activation of a specific semantic field in memory. A number of sophisticated approaches to examining the structure of this clustering have been developed, and a core theme is that the similarity relations between category members will reveal the mental semantic structure of the category underlying an individual’s responses, which can then be visualized by a number of algorithms, such as MDS, hierarchical clustering, ADDTREE, ADCLUS or SVD. Such approaches have been applied to a variety of neurological and psychiatric populations, and the general conclusion has been that the clinical condition systematically distorts the semantic structure in the patients, as compared to the healthy controls. In the present paper we explore this approach to understanding semantic structure using category fluency data. On the basis of a large pool of patients with schizophrenia (n=204) and healthy control participants (n=204), we find that the methods are problematic and unreliable to the extent that it is not possible to conclude that any putative difference reflects a systematic difference between the semantic representations in patients and controls. Moreover, taking into account the unreliability of the methods, we find that the most probable conclusion to be made is that no difference in underlying semantic representation exists. The consequences of these findings to understanding semantic structure, and the use of category fluency data, in cortical dysfunction are discussed.

Category fluency, latent semantic analysis and schizophrenia: a candidate gene approach

“…Specifically, the recall process likely involves a search for meanings as reflected in the ‘clustering’ of words in the output. Many approaches have been employed to examine the structure of the clustering, but are often problematic given the subjective judgments of cluster boundaries or have turned out to be simply unreliable (Voorspoels et al, 2013, in press). We have previously adopted LSA as an objective and reliable methodology to chart the flow of meaning in words and discourse (Elvevåg et al, 2007), and briefly describe this technique below.…”

Neuropsychological Characteristics of Verbal and Non-Verbal Fluency in Schizophrenia Patients

“…The study revealed that both factors independently affect verbal fluency function. A successive study failed to replicate the findings as it found no essential differences between schizophrenia patients and healthy controls in terms of cluster size, regardless of verbal fluency type (Voorspoels et al, 2014). Not only the number of clusters, but also their internal structure is analysed, which means examination of features such as distinction (how semantically different the clusters are), inclusion (a larger cluster can contain several smaller ones) and how overriding they are (parent cluster contains several smaller clusters).…”

“…Compared to healthy controls, results for schizophrenia patients showed differences in the number of generated words. However, the number of clusters, their quality and the frequency of switching between clusters or between clusters and words did not change (Voorspoels et al, 2014). Decreased speed of information processing was demonstrated to affect the number of generated words in the investigated group of patients (Hong et al, 2002;Riley et al, 2000;van Beilen et al, 2004;Vinogradov et al, 2003).…”

Clustering e switching em deficientes auditivos usuários do português brasileiro: fluência verbal semântica e fonológica

“…These findings were similar to the previous study by these researchers, in which there was no correlation between total evoked words and auditory characteristics of the subjects 1 . Another recent study hypothesized that although there may be quantitative differences in recalled words, the qualitative aspect of lexical organization does not change in different pathologies 26 .…”