## Evaluation of an Inference Network-Based Retrieval Model (1991)

Venue: | ACM Transactions on Information Systems |

Citations: | 229 - 20 self |

### BibTeX

@ARTICLE{Turtle91evaluationof,

author = {Howard Turtle and W. Bruce Croft},

title = {Evaluation of an Inference Network-Based Retrieval Model},

journal = {ACM Transactions on Information Systems},

year = {1991},

volume = {9},

pages = {187--222}

}

### Years of Citing Articles

### OpenURL

### Abstract

The use of inference networks to support document retrieval is introduced. A network-based retrieval model is described and compared to conventional probabilistic and Boolean models. The performance of a retrieval system based on the inference network model is evaluated and compared to performance with conventional retrieval models,

### Citations

1527 | Term-weighting approaches in automatic text retrieval
- Salton, Buckley
(Show Context)
Citation Context ...d in several ways, but previous information retrieval research has consistently shown within-document frequency ( tf ) and inverse document frequency ( idf ) to be useful components of such estimates =-=[32]-=-. In developing estimates, we concentrated on functions involving tf and idfi other functions are certainly possible and could be used in the basic model. The idf measure used in these experiments is ... |

618 | A Truth Maintenance System - Doyle - 1979 |

418 | The vocabulary problem in humansystem communication
- Furnas, Landauer, et al.
- 1987
(Show Context)
Citation Context ...Under this assumption, the same link matrix as for text nodes should be used. Research suggests, however, that the mismatch between query and indexing vocabularies may be a major cause of poor recall =-=[15]-=-. While our initial implementation is limited to linking query concepts to “nearly” equivalent representation concepts using a weighted-sum combination rule, it would appear that improved estimates of... |

417 | Probabilistic logic - Nilsson - 1986 |

316 | A generalization of Bayesian inference - Dempster - 1968 |

237 | Inference Networks for Document Retrieval
- Turtle
- 1991
(Show Context)
Citation Context ...improve retrieval performance when compared to equivalent networks with a single natural language query. Computational aspects of the model are discussed in [39] and complete results are presented in =-=[41]-=-. In the remainder of this section, we discuss techniques for estimating the required probabilities (Section 7. 1), compare performance with simple probabilistic searches (Section 7.2), and give resul... |

188 |
On relevance, probabilistic indexing and information retrieval
- Maron, Kuhns
- 1960
(Show Context)
Citation Context ... on the Bayesian inference network model. The use of Bayesian inference networks for information retrieval represents an extension of probability-based retrieval research dating from the early 1960’s =-=[23]-=-. It has long been recognized that some terms in a collection are more significant than others and that information about the distribution of terms in a collection can be used to improve retrieval per... |

166 |
Using Probabilistic Models of Document Retrieval without Relevance Information
- Croft, Harper
- 1997
(Show Context)
Citation Context ...esentations, but found that documents retrieved by multiple representations were likely to be relevant [20]. Similar results have been obtained when comparing term- with cluster-based representations =-=[2]-=- and termwith citation-based representations [16]. —Allow results from different queries and query types to be combined. Given a single natural language description of an information need, different s... |

113 |
Information retrieval by constrained spreading activation in semantic networks
- Cohen, Kjeldsen
- 1987
(Show Context)
Citation Context ...in information retrieval since at least the early 1960’s. Networks have been used to support diverse retrieval functions, including browsing [38], document clustering [7], spreading activation search =-=[4]-=-, support for multiple search strategies [11], and representation of user knowledge [27] or document content [40]. Recent work suggests that significant improvements in retrieval performance will requ... |

99 |
I3R: a new approach to the design of document retrieval systems
- Croft, Thompson
- 1987
(Show Context)
Citation Context ...early 1960’s. Networks have been used to support diverse retrieval functions, including browsing [38], document clustering [7], spreading activation search [4], support for multiple search strategies =-=[11]-=-, and representation of user knowledge [27] or document content [40]. Recent work suggests that significant improvements in retrieval performance will require techniques that, in some sense “understan... |

78 |
A definition of relevance for information retrieval
- Cooper
- 1971
(Show Context)
Citation Context ...ormation Systems, Vol. 9, No 3, July 1991, Pages 187-222.s188 . H. Turtle and W. B Croft The idea that retrieval is an inference or evidential reasoning process is not new. Cooper’s logical relevance =-=[6]-=- is based on deductive relationships between representations of documents and information needs. Wilson’s situational relevance [441 extends this notion to incorporate inductive or uncertain inference... |

76 |
Heuristic Reasoning about Uncertainty: An Artificial Intelligence Approach
- Cohen
- 1985
(Show Context)
Citation Context ... a representation node and let A, B, and C be document nodes. Let w., w~, and WCbe normalized tf values for A, B, and C, let idfq be a normalized idf weight for Q, and let t w~ = idfq” (w. + Wb+ WC). =-=(5)-=- Given our basic model, when A is instantiated, belief in Q is given by bel(Q) = ‘awq Wu+wb+wc — tf. . idf~(w. + Wb + WC) wa+w~+wc = tf. “ idf~ which is a form of tf. idf weight. In general, when a do... |

75 |
A Retrieval Model for Incorporating Hypertext Links
- Croft, Turtle
- 1989
(Show Context)
Citation Context ... that a user’s information need, expressed as one or more queries, is met given a document as “evidence.” Network representations show promise as mechanisms for inferring these kinds of relationships =-=[4, 12]-=-. This work was supported in part by the Air Force Office of Scientific Research under contract 90-0110 and by NSF grant IRI-8814790. Authors’ Address: Computer and Information Science Department, Uni... |

54 |
The Probability Ranking Principle
- Robertson
- 1997
(Show Context)
Citation Context ...olation because the approach is computationally simpler and because it allows comparison with earlier retrieval models that produce document rankings consistent with the Probability Ranking Principle =-=[29]-=- in which documents are considered in isolation. The document network is built once for a given collection. Given one or more queries representing an information need, we then build a query network th... |

54 |
Support for browsing in an intelligent text retrieval system
- Thompson, Croft
- 1989
(Show Context)
Citation Context ...ls. 1. INTRODUCTION Network representations have been used in information retrieval since at least the early 1960’s. Networks have been used to support diverse retrieval functions, including browsing =-=[38]-=-, document clustering [7], spreading activation search [4], support for multiple search strategies [11], and representation of user knowledge [27] or document content [40]. Recent work suggests that s... |

53 |
A study of the overlap among document representations
- Das-Gupta, Katzer
- 1983
(Show Context)
Citation Context ...same. Katzer, for example, found little overlap in documents retrieved using seven different representations, but found that documents retrieved by multiple representations were likely to be relevant =-=[20]-=-. Similar results have been obtained when comparing term- with cluster-based representations [2] and termwith citation-based representations [16]. —Allow results from different queries and query types... |

48 |
A model of cluster searching based on classification
- Croft
- 1980
(Show Context)
Citation Context ...k representations have been used in information retrieval since at least the early 1960’s. Networks have been used to support diverse retrieval functions, including browsing [38], document clustering =-=[7]-=-, spreading activation search [4], support for multiple search strategies [11], and representation of user knowledge [27] or document content [40]. Recent work suggests that significant improvements i... |

40 |
An inquiry into computer understanding
- Cheeseman
- 1988
(Show Context)
Citation Context ...is the familiar rule for disjunctive combination of events that are not known to be mutually exclusive. Similar matrix forms can be developed for and and not P(Q = true) = abc (2) P(Q= true) = 1 – a. =-=(3)-=- If we restrict the parent nodes for any of these logic operators to values O or 1 then Q must also have a value of O or 1. If we allow terms to take on weights in the range [0, 1] and interpret these... |

37 |
Information Retrieval. Butterworths
- Van-Rijsbergen
- 1979
(Show Context)
Citation Context ...ection 4.4); the estimation problems are essentially equivalent to those encountered with probabilistic or vector-space retrieval. 4.1 Probabilistic Retrieval ModeIs Conventional probabilistic models =-=[36, 42]-=- rank documents by the probability that each document would be judged relevant to a given query, P(relevant I dZ).2 This is, in many ways, similar to computing the probability that a user’s informatio... |

34 |
Approaches to intelligent information retrieval
- Croft
- 1987
(Show Context)
Citation Context ...27] or document content [40]. Recent work suggests that significant improvements in retrieval performance will require techniques that, in some sense “understand” the content of documents and queries =-=[9, 43]-=- and can be used to infer probable relationships between documents and queries. In this view, information retrieval is an inference or evidential reasoning process in which we estimate the probability... |

33 | Experiments with Query Acquisition and Use in Document Retrieval Systems
- Croft, Das
(Show Context)
Citation Context ...2) P(Q= true) = 1 – a. (3) If we restrict the parent nodes for any of these logic operators to values O or 1 then Q must also have a value of O or 1. If we allow terms to take on weights in the range =-=[0, 1]-=- and interpret these weights as the probability that the term has been assigned to a document text, then these inference networks provide a natural interpretation for Boolean retrieval with weighted i... |

33 |
An evaluation of factors affecting document ranking by information retrieval systems
- McGill, Koll, et al.
- 1979
(Show Context)
Citation Context ... of an information need, different searchers will formulate different queries to represent that need and will retrieve different documents, even when average performance is the same for each searcher =-=[20, 24]-=-. Again, documents retrieved by multiple searchers are more likely to be relevant. A description of an information need can be used to generate several query representations (e. g., probabilistic, Boo... |

29 | Characteristics of Two New Experimental Collections in Computer and Information Science Containing Textual and Bibliographic Concepts - Fox - 1983 |

26 |
Boolean queries and term dependencies in probabilistic retrieval models
- Croft
- 1986
(Show Context)
Citation Context ...(9) we have P(xl, . . ..xn) = IJJh) P(xlj . . . ~X.1fk) = IJJJ(~Ll fk) ‘(X,l fk) P(RldJ ‘p(Rlfk)p(fk)~ n P(x,ldm). Xsx lszsn P(xt) ACM Transactions on Information Systems, Vol 9, No 3, July 1991 4n). =-=(8)-=- and (8) (9) (lo)sEvaluation of an Inference Network-based Retrieval Model . 203 Fig. 5. Effect of inversion. The application of Bayes’ rule essentially inverts the network of Figure 4 to obtain the e... |

22 |
Coefficients of combining concept classes in a collection
- Fox, Nunn, et al.
- 1988
(Show Context)
Citation Context ...by multiple representations were likely to be relevant [20]. Similar results have been obtained when comparing term- with cluster-based representations [2] and termwith citation-based representations =-=[16]-=-. —Allow results from different queries and query types to be combined. Given a single natural language description of an information need, different searchers will formulate different queries to repr... |

13 | A Simple Blueprint for Automatic Boolean Query Processing - Salton - 1988 |

9 |
Experimental investigations of uncertainty in a rule-based system for information retrieval
- Tong, Shapiro
- 1985
(Show Context)
Citation Context ...unctions, including browsing [38], document clustering [7], spreading activation search [4], support for multiple search strategies [11], and representation of user knowledge [27] or document content =-=[40]-=-. Recent work suggests that significant improvements in retrieval performance will require techniques that, in some sense “understand” the content of documents and queries [9, 43] and can be used to i... |

7 | The Effect of Document Ranking on Retrieval System Performance: A Search for an Optimal Ranking Rule - Stirling - 1977 |

6 |
The use of adaptive mechanisms for selection of search strategies in document retrieval systems
- Croft, Thompson
- 1984
(Show Context)
Citation Context ...erally rank a relatively small number of documents highly, after which belief values drop off rapidly. Or-structured queries tend to produce a more uniform distribution of beliefs. Croft and Thompson =-=[10]-=- found that different query representations or strategies worked better for some queries than others and that it was difficult to predict which strategy would work best with a given query. In their wo... |

4 |
editors. Uncertainty in Artificial Intelligence
- Kanal, Lemmer
- 1986
(Show Context)
Citation Context ...inds of inference are similar to those used in expert systems that must reason with uncertain information. A number of competing inference models have been developed for these kinds of expert systems =-=[19, 21]-=- and several of these models can be adapted to the document retrieval task. In the research described here we adapt an inference network model to the retrieval task. The use of the model is intended t... |

4 |
Representation of anomalous states of knowledge in information retrieval
- Oddy, Palmquist, et al.
- 1986
(Show Context)
Citation Context ...pport diverse retrieval functions, including browsing [38], document clustering [7], spreading activation search [4], support for multiple search strategies [11], and representation of user knowledge =-=[27]-=- or document content [40]. Recent work suggests that significant improvements in retrieval performance will require techniques that, in some sense “understand” the content of documents and queries [9,... |

3 | Research on automatic indexing 1974-1976 - K, Bates - 1977 |

3 | A Mathematid Theory of Euidence - Shafer - 1976 |

2 |
Automatic ranked output for Boolean searches in SIRE
- Noreault, Koll, et al.
- 1977
(Show Context)
Citation Context ...ression evaluation. It is also possible that some form of idf weighting could be used to improve performance of Boolean queries or that weighting strategies developed for conventional Boolean systems =-=[26]-=- could be adapted to the network model. An experiment in which each intermediate and and or expression was weighted using an idf based on the number of documents in the intermediate result did not imp... |

2 |
A mm-classical logic for information retrieval
- RIJfiDF, J
- 1986
(Show Context)
Citation Context ...27] or document content [40]. Recent work suggests that significant improvements in retrieval performance will require techniques that, in some sense “understand” the content of documents and queries =-=[9, 43]-=- and can be used to infer probable relationships between documents and queries. In this view, information retrieval is an inference or evidential reasoning process in which we estimate the probability... |

1 |
Models for retrieval with probabilistic indexing
- FURH
- 1989
(Show Context)
Citation Context ... with that shown in Figure 1, we ‘Most probabilistic models do not actually compute P(relevant I d,), but simply rank documents using some function that is monotonic with P(relevant I cl,). Like Fuhr =-=[18]-=-, we believe that an estimate of the probability of relevance is more useful than the ranking by itself A ranked hst of documents in which the top ranked document has a probability of relevance of O,5... |

1 |
Local computations with probabihties on graphical structures and their application to expert systems
- LAURITZEN, SPIEGELHALTER
- 1988
(Show Context)
Citation Context ...y symbolic reasoning [5, 141, fuzzy sets [45], and a variety of probability models [3, 251. Two inference models based on probabilistic methods are of particular interest: Bayesian inference networks =-=[22, 28]-=- and the Dempster - Shafer theory of evidence [13, 341. A Bayesian inference network is a directed, acyclic dependency graph (DAG) in which nodes represent propositional variables or constants and edg... |

1 |
Probabilistic Reasoning in Intelhgent Systems: Networks of PlausLble Inference
- PEARL
- 1988
(Show Context)
Citation Context ...y symbolic reasoning [5, 141, fuzzy sets [45], and a variety of probability models [3, 251. Two inference models based on probabilistic methods are of particular interest: Bayesian inference networks =-=[22, 28]-=- and the Dempster - Shafer theory of evidence [13, 341. A Bayesian inference network is a directed, acyclic dependency graph (DAG) in which nodes represent propositional variables or constants and edg... |

1 |
Extended Boolean mformatlon retrieval
- SALTON, Fox, et al.
- 1983
(Show Context)
Citation Context ...ecision at three recall levels (25%, 50%, and ‘7570) rather than our customary ten levels, in order to permit comparison with the published results. Note that, for the CACM collection, Salton, et al. =-=[33]-=- include two queries (author searches) that are not used in our experiments. For both CACM and CISI, the network evaluation of the Boolean queries is better than the best p-norm evaluation. For CACM, ... |

1 |
Non-parametric Statist~cs for the Behavorial Sczences
- SIEGEL
- 1956
(Show Context)
Citation Context ...precision at ten standard recall points. When two tests are being compared, we show the difference as the percent change from the baseline test. Significance tests are based on a one-tailed Sign test =-=[35]-=- comparing the ten averaged precision values for each query set, where a 5 percent difference in average precision is required for two observations to be considered different. This is a conservative t... |

1 |
Introduction to Modern Information RetrLeval
- MCGILL
- 1983
(Show Context)
Citation Context ...ditional information to the query network or refine its structure to better characterize the meaning of the existing queries. This feedback process is quite similar to conventional relevance feedback =-=[36]-=-. 3.4 Link Matrix Forms For all non-root nodes in the inference network we must estimate the probability that a node takes on a value given any set of values for its parent nodes. If a node a has a se... |

1 |
Efficient evaluation for probabilistic retrieval
- TURTLE, CROFT
(Show Context)
Citation Context ...s and search strategies will significantly improve retrieval performance when compared to equivalent networks with a single natural language query. Computational aspects of the model are discussed in =-=[39]-=- and complete results are presented in [41]. In the remainder of this section, we discuss techniques for estimating the required probabilities (Section 7. 1), compare performance with simple probabili... |

1 | Storage RetrLeval 9 - WILSON, relevance, et al. - 1973 |