SP500215 TREC-II Routing Experiments with the TRW/Paracel Fast Data Finder chapter M. Mettler National Institute of Standards and Technology D. K. Harman reports, etc.) are not relevant to this topic; nor are articles which describe subsidies not directed toward Airbus. Traditional IR term weighting techniques do not give any explicit benefit to articles which conjoin ideas. Articles which include terms relevant to each of the component sub-topics will receive high scores; but so will articles which include many terms relevant to only one sub-topic. Recent efforts implicitly include conjunctions through the use of phrases as terms in otherwise traditional statistical methods. An alternative is the use of boolean operators. This has the desired effect -- an AND of terms forces a conjunction -- but the use of booleans in IR has been viewed with some skepticism and disfavor. Boolean operators often find a conjunction of terms where none truly exists (for example, Airbus and subsidies might be mentioned in two separate and unrelated portions of an article); or, if made sufficiently restrictive to eliminate spurious matches, boolean-based searches often miss relevant articles. We have followed an approach which incorporates both ideas. Rather than focus on specific phrases, we search for terms in proximity to one another. The terms in the query are chosen to represent each of the constituent sub-topics, just as in a boolean search. The specificity of the query is adjusted by varying the required proximity of the terms. Thus, for Airbus subsidies we might search for terms representing "Airbus" in a range of proximities to terms representing "subsidies". This approach allows conjunctions to be graded. A small proximity restriction (say, 3 words) yields results similar to a keyphrase search, indicating that the two concepts are indeed associated in the article and that the article is relevant to the topic. A large proximity restriction (1 article) is analogous to a simple boolean keyword search and retreives articles in which the concept terms may be only loosely associated. Intermediate proximities (1 sentence, 1 paragraph, etc.) indicate intermediate degrees of association and intermediate recall/precision trade-offs. It is also possible to use multiple proximities in a single query with this method, or to use proximities and occurrence frequencies together, to form multi-dimensional arrays of query parameters. For example, for Topic 62, Military Coups Dtetat, the number of conjunctions was traded off against the proximity of the conjunction to form a two-dimensional query set. For the initial experiment, lists of synonyms representative of each idea in a topic were manually built, and one- or two-dimensional query sets were built from these lists. These queries were then run against the training database, and after some feedback, the query sets were finalized. Each finalized query set was run against the training database to determine a ranking of the queries based solely on selectivity. Table I shows a sample proximity query and Table III shows our TREC-Il results. The number of relevant documents retrieved by the proximity method queries are labeled TRW1. 203