The Intelligent Ignoramus, an Experiment in Numerical Taxonomy

Summary The feasibility of using technicians untrained in taxonomy to collect data for use in numerical taxonomic studies was investigated. Two persons independently described twenty‐five species of bees (given 6 specimens of each species, 3 of each sex). The results of the analysis of their descriptions were compared with the data of Michener and Sokal (1957). The data were processed using standard numerical taxonomic procedures (standardization of characters, computation of correlation and distance coefficients among species, cluster analysis, principal components analysis and 3‐D models). The 3‐D models revealed that the generic groupings were not nearly as distinct in the data obtained from the technicians as they were in the original data collected by Michener and Sokal, ( M&S , used as a standard for comparison). When the data collected by both technicians were pooled, they yielded much better agreement with the M&S data than either one did by itself. A similar conclusion was derived from examination of the phenograms. While the phenograms made good sense overall, there were many discrepancies between those of the technicians and those of M&S An analysis of the correlations among the 234 characters from both technicians as well as from M&S enabled a preliminary check for the matches asymptote hypothesis to be made. A cluster analysis of the character correlation matrix revealed that there were no characters which were obvious duplicates (no correlations equal to unity). The same morphological structures were examined but they were described in different ways. Rotation to an oblique simple structure of 17 factors obtained by factor analysis indicated that the character suites described in the three studies shared dimensions of variability. One of the technicians had not described characters loading highly on three of the factors and the M&S data lacked characters loading highly on one of the factors. An important feature of the data is the large number of unique characters (reflecting independent dimensions of variation). These ranged from 18% to 40%. We suspect that these unique characters indicate the presence of taxonomic “noise”. The results of this study would seem to imply that characters defined by untrained technicians could lead to acceptable classifications if enough characters are considered. Consequently, we are encouraged about prospects for the eventual automation of the data gathering process in taxonomy.