Relational thesauri in information retrieval

Abstract This article describes the design and development of a new type of thesaurus based on lexical‐semantic relations. Relational thesauri have been constructed to perform a new type of term classification. These relational thesauri are generally applicable to any document collection and their maintenance is relatively simple. A series of experiments to evaluate thesauri of this new type have been run on an information retrieval system called IRS at Illinois Institute of Technology. The results of experiments with queries enhanced using thesauri based on several different groups of relations have been compared against performance with the original queries. Thesauri from most groups, except antonyms, made improvements in recall as well as in precision. The best results come from a set of ill‐formed queries with few index terms. These results have been analyzed with both precision‐recall graphs and statistical tests. While thesauri that combine a number of relations together have been most effective in a batch environment, there is reason to believe that individual relations will be more useful in an interactive retrieval system that presents index terms to the user and allows him to choose those that best convey his meaning.