ISR11 Scientific Report No. ISR-11 Information Storage and Retrieval Relevance Feedback in an Information Retrieval System chapter W. Riddle T. Horwitz R. Dietz Harvard University Gerard Salton Use, reproduction, or publication, in whole or in part, is permitted for any purpose of the United States Government. VI-lo E) Termination of the Modification Process Updating is terminated when all relevant documents have been retrieved (since the user's needs are then satisfied as fully as possible), or when at most three modifications have been made (since the results presented in Figure 6 indicate that, with the progressions of [OCRerr] used in the final investigations, the iteration process can safely be terminated in general after three modifications have been made). 3. I[OCRerr]xperimental Results In general, the modification of a query using relevance feedback information leads to an improvanent in both the ni[OCRerr][OCRerr]ber of relevant docu- ments retrieved and in the ranks of all the relevant documents. The modi- fication normally yields an increase in both precision and recall (as * shown in Figure 2), regardless of how a is applied, provided that the set of relevant documents lies in one basic cluster in n-space. If the relevant documents cluster in two separate regions in n-space (as a result of the indexing scheme used), the results are as shown in Figure 7. When such a dual clustering of the relevant documents exists, Rocchio suggests the use of [OCRerr]tiple queries. [2) This is good theoretically, when a priori relevance judgments, which list all the documents relevant to a given query, have been made. However, in a real system, the user is uncertain of the existence of other relevant documents and the technique is impossible to carry out. A possible solution is the use of a list that guarantees, for example, that whenever document X, Y, and Z are deemed relevant, then documents A, B, and C are also relevant and are returned * The statement that the iterative retrieval process does not significantly depend on the particular strategy of applying a (for the progressions of a used in the final investigations) is supported by the data given in Figure 6, for the progressions used and the correlation function.