SP500207 Overview of the First Text REtrieval Conference (TREC-1) chapter D. Harman National Institute of Standards and Technology Donna K. Harman Figure 12 shows the recalVprecision curve for the three TREC-1 runs with the highest 1 1-point averages using manual construction of queries. The systems used manually-built filters, with the "clartb" and `tgecrd2" results done similarly to their corr[OCRerr]sponding adhoc systems, but using the sample relevant documents as input to the filter-building process. The "parazi" system used manually-constructed filters based on clusters of interest- mg terms (see Zimmerman paper). The "cpghc2" group hand-crafted these queries as a contrast to their automatic pattern filtering methods (see Jones, Leung & Pape). Best Manual Routing 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00 Recall parazi gecrdl gecrd2 trwl 0.9 0.8 0.7 [OCRerr] 0.6 0 [OCRerr] 0.5 [OCRerr] 0.4 0.3 0.2 0.1 0 Figure 12. The Best Routing Results using Manual Query Construction. Again it is useful to contrast the methods of query construction. Figure 13 shows a comparison of four sets of results, two from automatic query construction and two using manual query construction. Here, unlike the adhoc results, the automatic query building seems to be clearly superior, with the "fuhrl" results having higher performance throughout the significant part of the recalVprecision curve. Routing Manual vs Automatic 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1*00 Recall 0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 [OCRerr]- ftihrl __ cnil[OCRerr] . parazi gecrd2 Figure 13. A Comparison of Routing Results using Different Query Construction Methods. 18