A visualization case study of feature vector and stemmer effects on TREC topic-document subsets

A method of visual analysis is demonstrated which takes advantage of the "p ooling" technique of topic-document set creation in the TREC collection. TR EC topic-document sets create a specific pattern when converted to similari ty measures, scaled, and plotted. Using the visual pattern created by full text as a normative view of the data, the effect of feature vectors and ste mming on recovering the normative view is shown visually. When stemmed, fea ture vectors of length 200 were shown to substantially recover the normativ e visual configuration created by full text. Some caution regarding the use of stemming is indicated by the dispersion of documents in the visual fiel d if feature vector approaches are to be applied to filtering tasks.