Document Retrieval and Routing Using the INQUERY System

The INQUERY retrieval and routing system, which is based on the Bayesian inference net retrieval model, has been described in a number of papers 5, 4, 10, 11]. In the TREC experiments this year, a number of new techniques were introduced for both the ad-hoc retrieval and routing runs. In addition, experiments with Spanish retrieval were carried out. For the ad-hoc retrieval experiments, the major changes to the system were the incorporation of passage retrieval, query expansion using PhraseFinder 8], a new estimation technique for indexing probabilities, and new analysis techniques for the TREC topics. The following description of query processing emphasizes the diierences to the approach used in the previous evaluations, which is described in 5, 3]. The text in all parts of the TREC topics (topic, description, narrative) were treated in the same way. The University of Massachusetts JTAG tagger 12] was used to identify parts of speech for words in the query. Then sequences of nouns, or sequences of adjectives and following nouns, were selected for the #PHRASE operator. (In the past, we had included prepositional phrases which followed a \bare" headnoun. We eliminated these phrases because we found they were not useful.) Analysis of phrase statistics has shown that phrases of length greater than two words were invariably too restrictive, especially when subphrases of these would have been useful in retrieval. We therefore used only two-word phrases. Sequences of more than two words were broken down into two-word subsequences. For example: (crude oil price trend)) (crude oil), (oil price), (price trend). Using this modiication of the original phrase extraction procedures, we were able to eliminate a number of special processing steps, and apply noun-phrase-extraction processing to all sections.