Research in the probabilistic theory of information retrieval involves the construction of mathematical models based on statistical assumptions.One of the hazards inherent in this kind of theory construction is that the assumptions laid down maybe inconsmtent in unanticipated ways with the data to which they are applied.Another hazard is that the stated assumptions may not be those on which the derived modeling equations or resulting experiments are actually based.

The goal of a probabilistic retrieval system design is to rank the elements of the search universe in descending order of their estimated probability of usefulness to the user. Previously explored methods for computing such a ranking have involved the use of statistical independence assumptions and multiple regression analysis on a learning sample. In this paper these techniques are recombined in a new way to achieve greater accuracy of probabilistic estimate without undue additional computational complexity. The novel element of the proposed design is that the regression analysis be carried out in two or more levels or stages. Such an approach allows composite or grouped retrieval clues to be analyzed in an orderly manner -- first within groups, and then between. It compensates automatically for systematic biases introduced by the statistical simplifying assumptions, and gives rise to search algorithms of reasonable computational efficiency.

Article Some inconsistencies and misnomers in probabilistic information retrieval Share on Author: William S. Cooper School of Library and Information Studies, University of California, Berkeley, California School of Library and Information Studies, University of California, Berkeley, CaliforniaView Profile Authors Info & Claims SIGIR '91: Proceedings of the 14th annual international ACM SIGIR conference on Research and development in information retrievalSeptember 1991 Pages 57–61https://doi.org/10.1145/122860.122866Online:01 September 1991Publication History 31citation267DownloadsMetricsTotal Citations31Total Downloads267Last 12 Months6Last 6 weeks1 Get Citation AlertsNew Citation Alert added!This alert has been successfully added and will be sent to:You will be notified whenever a record that you have chosen has been cited.To manage your alert preferences, click on the button below.Manage my AlertsNew Citation Alert!Please log in to your account Save to BinderSave to BinderCreate a New BinderNameCancelCreateExport CitationPublisher SiteGet Access

Abstract Several of the drawbacks of conventional information retrieval systems can be overcome by a design approach in which queries consist of sets of terms, either unweighted or weighted with subjective term precision estimates, and retrieval outputs are ranked by probability of usefulness estimated in accordance with the so‐called “maximum entropy principle.” A system organized along these lines combines the convenience of a simple input language with a powerful probabilistic inference mechanism capable of exploiting kinds of statistical clues not ordinarily used in systems of traditional design. The sensitivity of the maximum entropy principle to the frequencies and joint frequencies with which terms have been assigned to documents in the collection results in a system design of increased power and expressiveness without a concomitant increase in the complexity of the request language. It incorporates the more important search capabilities of both Boolean and conventional weighted‐request languages and facilitates the use of unconventional search clues.

One of the most perplexing problems of reformation retrieval has been the estabhshment of rational criteria for deciding what index terms or descriptors to assign to a unit of stored information for purposes of later retrieval Both probablhstJc and utlhty-theoretlc criteria have m the past been proposed for thts purpose.The present paper derives explicit decision rules of both kinds from a common conceptual and mathematical foundation The result IS a unified theory of indexing KEY WORDS AND PHRASES indexing, cataloging,

Abstract Although it is widely accepted that indexing and cataloging are central to the document and reference retrieval process, there is as yet little unanimity on the issue of how these operations ought to be carried out. Since indexing is a decision‐making process, decision theory and utility theory have been proposed as a possible theoretical framework within which to formulate decision rules for indexing. These decision rules may be interpreted operationally as guides to the thought processes appropriate to the making of sound indexing decisions. One such thought process involves “gedanken experimentation,” in which the indexer is led to an indexing decision by performing suitable thought experiments. This approach to indexing, in addition to providing a theoretical basis for what the indexer does, lends itself to various training schemes for the education of indexers, to the provision of graphic aids to indexing, and to the evaluation of indexing performance. There are also ramifications for automatic indexing.

Abstract It is argued that a user's subjective evaluation of the personal utility of a retrieval system's output to him, if it could be properly quantified, would be a near‐ideal measure of retrieval effectiveness. A hypothetical methodology is presented for measuring this utility by means of an elicitation procedure. Because the hypothetical methodology is impractical, compromise methods are outlined and their underlying simplifying assumptions are discussed. The more plausible the simplifying assumptions on which a performance measure is based, the better the measure. This, along with evidence gleaned from ‘validation experiments’ of a certain kind, is suggsted as a criterion for selecting or deriving the best measure of effectiveness to use under given test conditions.

Abstract It was argued in Part I (see JASIS, March‐April 1973 p. 87) that the best way to evaluate a retrieval system is, in principle at least, to elicit subjective estimates of the system's utility to its users, quantified in terms of the numbers of utiles (e.g. dollars) they would have been willing to give up in exchange for the privilege of using the system; and a naive methodology was outlined for evaluating retrieval systems on this basis. But the impracticality of the naive evaluation procedure as it stands raises the questions: How can one decide which practical measure is likely to yield results most closely resembling those of the naive methodology? And how can one tell whether the resemblance is close enough to make applying the measure worth while? In the present paper two kinds of solution to these problems are taken up. The first answers the questions in terms of the reasonableness of the simplifying assumptions needed to get from the naive measure to the proposed substitute. The second answers it by experimentation.

Information Retrieval systems may be classified either as Document Retrieval systems or Fact Retrieval systems. It is contended that at least some of the latter will require the capability for performing logical deductions among natural language sentences. The problem of developing systems of logical inference for natural languages is discussed, and an example of such an analysis of a sublanguage of English is presented. An experimental Fact Retrieval system which incorporates this analysis has been programmed for the IBM 7090 computer, and its main algorithms are stated.