Combining evidence for effective information filtering

As part of NLqT/ARPA’s TREC Wcckshop, we used Semamk Indexing (LSl) for filtering 33~ incoming dotemere from diveae ~ (aew~es, I~ttenta tedmiotl ~mractO for SO t~i~ of ~ We develot~edrepmeematiom of user intems~ ~ mtea, for these topics ~ two mu~ces of emmi~ infmmalion. A Won/F/~r used just the weeds in the ~#c mmnems, and a Re/Do~ F//~r u~d just ̄ e known n/ram t~s do~.me~ and ~ ~e ~#c ,memm. U~ the mlevam mdnin8 documents (a variant of relevance &edback) was mote effective than ~ ̄ detailed natural ’-,,-mage description of in~ts. Coral/nins these two vec~n provided some ~ iml~ovemmm in mmdn~ On averap, 7 ~ the top 10 docmnmes ~md 44 of the top 1(]0 ~ ~ mlev ̄llt nslno the combined ~ metlwd. Dam coml/nation ’~n~_ the results of the Wa/d and RelDoca reuieval .-ts was not generally succe,dul in ~ pe~onm~e ~oqmed m the best individual method, thoNlh we believe it might be if additional murces me used. These ~ mmtmds me quite gen. end and applicable to ̄ variety of muting aad feedback