SP500207 Site Report for the Text REtrieval Conference chapter P. Nelson National Institute of Standards and Technology Donna K. Harman The results show that ConQuest was best on five of the 39 queries, worst on one query, and well above average for most. EssentiallY, the area above average is much greater than the area below. The next chart (figure 6) compares ConQuest to the next two highest Category A manual systems. This chart was originally produced in the TREC proceedings. 1.0000 0.9000 0.8000 , 0.7000 Cu -[OCRerr] 0.6000 0 0.5000 0.4000 0 ; 0.3000 e[OCRerr] 0.2000 0.1000 0.0000 0 o 0 0 0 0 0 0 0 0 o Lc[OCRerr] o 0 0 0 0 0 0 0 0 0 [OCRerr]ConQuest - . - Recall percentage Figure 6 precision at Recall for the Top 3 Scorers Category A, Manual Mode This chart shows ConQuest having be[OCRerr]r perf0rmance in the high-recall region (where a large percentage of the relevant documents are retrieved). We suspect that the performance in this region is enhanced by our aggressive expansion of terms coupled with the flexible ranking algorithm. Good performance in the high recall region is important to us, because studies have shown that high recall is the most difficult problem for text retrieval systems. Many systems achieve relatively high precisiofl by discarding documents, but very few achieve high recall. 295