Abstract
A Web image search application was built using a previously-developed image relevance model for retrieval of images via text-based image retrieval. The application includes a text stemmer that converts a word to a canonical form, making it possible to match text in the face of changes in tense or plurality that have little effect on semantics. The usefulness of stemming in Web image retrieval was evaluated via a test en ten queries that were submitted both with and without stemming. Relevance of retrieved images was determined via ratings by three trained individuals. With stemming, the average unique relevance recall (a measure of the proportion of relevant images returned by one algorithm and not another) was 27.7%, while without stemming, it was only 0.5%. These results may more accurately apply to queries containing at least one plural noun, present tense verb, present participle verb, or past tense verb.