IRS13 Scientific Report No. IRS-13 Information Storage and Retrieval Suffix Dictionaries chapter E. M. Keen Harvard University Gerard Salton Use, reproduction, or publication, in whole or in part, is permitted for any purpose of the United States Government. VI-l7 This data shows that full suffixing (stem dictionary) does not affect so many aerodynamics words as it does in computer science and documentation, thus giving the Cran-l collection less of an opportunity for a change in retrieval performance. Further explanations can be given by observing in- dividual request performance for the seven requests that have performance changes greater than 0e05 no[OCRerr][OCRerr]ized recall (see Figure 6); tour of these requests are better on stem, and three are better on suffix `5'. An analysis of the three requests that favor suffix [OCRerr] reveals certain test problems, connected mainly with hyphenation and keypunch errors, and in fact the request/ relevant document match is in all cases very week. The many requests that favor suffix `5' by a trivial amount (Figure 6) are typified by request [OCRerr]269, details of which appear in Figure Il. As Figure U shows, the stem compress'1 incorrectly matches the request word 11compressor" with a frequently used word "compressible'1 in two non-relevant documents, so that the stem dictionary has an interior performance because relevant document 1590 receives a rank position below the two non-relevant documents. In case the matching [OCRerr]rds of non-relevant document l98i[OCRerr] appear to put into question the relevance decision (the abstract includes the topic of 11choked flow in an impeller inlet 11), the title makes it clear that it is a "centrifugal impeller", and the matching word "axial" is a qurious match frcm the phrase "axial [OCRerr][OCRerr]etry'1. The ranks of the seven documents relevant to request Ql90 are given in Figure 12, because the changes in rank position observed are typical of what happens to the averages. The highest ranked relevant document remains unchanged at position 1; thus the high precision end of the curve for this request will remain unchanged. The next three relevant items are ranked better by suffix `5', but the final three relevant are ranked better by stem. This result is seen to give a greater superiority to suffix `8' in