SP500215 The ConQuest System chapter P. Nelson National Institute of Standards and Technology D. K. Harman simple two or three word phrase. Therefore, a fine-grain rank tuned for short queries has provided the best and most accurate system for our commercial users. In TREC-2, however, the queries are often 40-50 words long. These analyses have shown us that our initial fine- grain algorithms are not as accurate for such long queries. Future Analysis Our tests indicate that more study of other fine-grain algorithms is where ConQuest can most likely improve its scores for TREC-3. New ways of looking at proximity and positional information in the document will be explored and compared against the existing coarse-grain ranking results. We still feel that our existing fine-grain algorithm is best for the typical commercial user, and we are looking for ways to fully test this hypothesis. Finally, we are now much more sensitive to the effects of query size on fine-grain algorithms and are looking more closely at ways to desensitize our fine-grain algorithms, or to adapt them easily to different query lengths. 270