, <desc>, and <narr>) with respect to the training database (disk 1). They were subsequently run in the routing mode against the test database (disk 2). 26 Unfor- tunately, the routing results were hindered by a bug in our program that removed duplicate 26 While other fields provided much useful information about various aspects of a given topic (including definitions of spe- cialized terms), we thought that their inclusion in the final query would make it difficult to assess the advantages of linguistic pro- cessing. This especially applies to Concepts fiels <con> which list- ed hand-extracted keywords. Subsequenfly, we realized that more often this field in fact provided further clues not found in other sec- tions of a topic.

wordi word2 SIMnorm abm absence accept accord acquire speech adjustable maxsaver affair affordable disease medium+range aircraft aircraft airline alien anniversary anti+age anti+clot contra candidate contend property attempt await stealth child baggage ban bearish bee roller+coast two +income television soldier treasury research withdrawal Table 2. Filtered more specific term). *anti+ballistic *maternity acquire pact purchase address one +year *advance +purchase scandal low+income *ailment *air+to+air *jetliner plane carrier immigrate *bicentennial anti+wrinkle cholesterol+lower *anti+sandinista *aspirant *aspirant asset bid pend *b+1 *baby luggage restrict bullish *honeybee *bumpy two+earner tv troop *short+term study *pullout 0.534894 0.233082 0.179078 0A92332 0A49362 0.263789 0.824053 0.734008 0.684877 0.181795 0.247382 0.874508 0.166777 0.423831 0.345490 0.270412 0.588210 0.153918 0.856712 0.294677 0.116025 0.143459 0.285299 0.641592 0.572960 0.877582 0.183064 0.607333 0.321943 0.847103 0.461023 0.898278 0.293104 0.806018 0.374410 0.661133 0.209257 0.622558 word similarifies (* indicates the terproductive with larger databases. Later experiment have confirmed this as we obtained new code from NIST. 183 wordi word2 SIMnorm lord abuser 0.261507 break accelerator 0.106775 accept reject 0.275802 accord level 0.152023 cent accrue 0.259478 acquire deficient 0.615525 fix+rate adjustable 0.930180 advertise environmental 0.124026 petroleum aerospace 0.207406 afghan nicaraguan 0A21478 banana dominican 0.444193 begin month 0.346558 german british 0.112465 superpower reagan+ gorbachev 0.400145 republic sea 0.227662 sunglasses sneakers 0.126749 Table 3. Some (apparently?) "bad" similarities gen- erated. also quite disappointed with the effect that the query expansion had (or did not have) on either precision or recall figures. This was somewhat more directly related to the wide domain scope of WSJ articles, and we noticed many obvious word sense mixups leading to unwanted term associations. Also, the rather straightforward method of query expansion terms proved less adequate here, and we are considering new methods as outlined briefly in the introduction to this paper. Our final test with manually altered queries (terms were deleted or added after consulting the original topic) brought only negative results: any kind of intervention appeared only to further decrease both recall and precision. Final runs were performed as follows: (1) Topics 1-25 were processed using only direct fields (, <desc>, and <narr>) with respect to the training database (disk 1). They were subsequently run in the routing mode against the test database (disk 2). 26 Unfor- tunately, the routing results were hindered by a bug in our program that removed duplicate 26 While other fields provided much useful information about various aspects of a given topic (including definitions of spe- cialized terms), we thought that their inclusion in the final query would make it difficult to assess the advantages of linguistic pro- cessing. This especially applies to Concepts fiels <con> which list- ed hand-extracted keywords. Subsequenfly, we realized that more often this field in fact provided further clues not found in other sec- tions of a topic.