Application of Logical Analysis of Data to the TREC-6 Routing Task

In the official runs, statistics are produced using the MG (Witten, Moffat, Bell, 1994)) search engine, and the terms are in fact stems, rather than complete terms. The top 25 terms selected from the positive and negative examples are merged, to form a list with no more than 50 terms. The MG retrieval system is used (massively) to transform every judged document into a Boolean vector with one component for each distinct classification term. The RUTCOR LAD program (Boros, Hammer, Ibaraki, Kogan, Mayoraz, & Muchnik, 1996) is used (twice for each topic), with several modifications, to search exhaustively for Boolean prime implicants which characterize the positive and the negative examples. Due to computer speed limitations, we have limited the search in our official submissions to terms of order three (i.e terms such as ABC', where C' denotes the absence of term C). Each pattern which matches some positive (respectively, negative) examples is given a weight determined by the number of examples that it matches.