SP500215 TREC-2 Routing and Ad-Hoc Retrieval Evaluation using the INQUERY System chapter W. Croft J. Callan J. Broglio National Institute of Standards and Technology D. K. Harman experiments, we have developed a new stemming algorithm that has a number of advantages for operational systems. A number of recognizers written in flex are then used to identify objects such as company names and mark their presence in the document using "meta" index terms. A company name such as IBM in the text, for example, will result in a meta term #COMPANY being recorded at that position in the text. The use of these meta terms extends the range of queries that can be specified. This completes the usual processing for document text. The document indexing process also involves building the compressed inverted ifies that are necessary for efficient performance with very large databases. Since positional information is stored, overhead rates are typically about 40% of the original database size. The query processing process involves a series of steps to identify the important concepts and structure describing a user's information need. INQUERY is unique in that it can represent and use complex structured descriptions in a probabilistic framework. Many of the steps in query processing are the same as those done in document indexing. In addition, a part-of-speech tagge? is to used to identify candidate search phrases. Domain-dependent features are recognized and meta-terms inserted into the query representation. The relative importance of query concepts is also estimated, and relationships between concepts are suggested based on simple grammar rules. An evaluation of some of the query processing techniques is presented in [1]. INQUERY also has the capability of expanding the query using relationships between concepts found by either using manually specified domain knowledge in the form of a simple thesaurus or by corpus analysis. The WORDFINDER system is a version of INQUERY that retrieves concepts that are related to the query. WORDFINDER is constructed by identifying noun groups in the text and representing them by the words that are closely associated with them (i.e. occur in the same text windows). Concept "documents" are then stored in INQUERY. This technique of query expansion was not tested in TREC-2. The query evaluation process uses the inverted ifies and the query represented as an inference net to produce a document ranking. The evaluation involves probabilistic inference based on the operators defined in the INQUERY language. These operators define new concepts and how to calculate the belief in those concepts using linguistic and statistical evidence. We are constantly experimenting with and refining these operators (for example, the operator defining a phrase-based concept) in order to improve retrieval performance. The relevance feedback process uses information from user evaluations of retrieved doc- uments to modify the original query in detection or routing environments. The INQUERY system, because it can represent structured queries, supports a wide range of learning tech- niques for query modification [5]. In general, new words and phrases are identified in the sample of relevant documents. These are added to the original query and all the terms in the query are then reweighted. With the amount of relevance information available in TIPSTER, relatively simple automatic techniques appear to produce good levels of effec- tiveness. We are also investigating the effect of using more limited information and more complex learning techniques, such as neural networks. 77