H. Turtle and W. B. Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187--222. ACM, July 1991.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....by the input system, and use the entire document as input to the metasearch algorithm. Ng and Kantor call this data level fusion. Vogt s ACE model [74] explores this approach in an experimental framework, and the NECI metasearch engine [39] implements it on the Web. The INQUERY retrieval system [72] can be considered another example. In this work, we restrict ourselves to the two most common cases, when either only ranks or ranks plus relevance scores are available. See Figure 1.1. Training data. Another dimension in Figure 1.1 is the presence or absence of training data. Training data ....

....the best performance was achieved by using both information sources. 29 Other studies in the 1970s and 80s [50, 36, 64] showed that using di erent query representations or di erent document representations led to very di erent retrieval results. Data Level Fusion In 1991, Turtle and Croft [72] proposed the the INQUERY retrieval system, based on inference networks. The inference network model is a general model for combining evidence; INQUERY falls in the category of data level fusion, where the documents and queries themselves are available to the fusion algorithm, and thus is outside ....

Howard Turtle and W. Bruce Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187-222, July 1991.

....contained in the indexing terms their choice and generation is the crucial step in handling information semantics. Results of experiments have shown that document retrieval using stemmed natural language terms taken from a document for indexing it is comparable to the use of controlled languages [Turtle and Croft, 1991]. However, it is argued that the use of compound expressions or propositional statements (very similar to RDF) will increase precision and recall [Lewis, 1996] The crucial task in using natural language as a source of semantic information is the analysis of documents and the generation of ....

Turtle, H. and Croft, W. (1991). Evaluation of inference network-based retrieval methods. ACM Transactions on Information Systems, 9(3):187--222.

....about multiple sources of uncertain evidence, which makes them particularly suited for IR. Previous research showed that INQUERY s performance, as measured by precision and recall, generally improved with the amount of evidence available. This result held for collections of short documents [16], a collection of large documents [7] and a heterogeneous collection [1] INQUERY s effectiveness at using multiple sources of evidence is consistent with the view of IR as a task of retrieving structured documents. Multiple sources of evidence can be obtained by applying a query to various ....

H. R. Turtle and W. B. Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187-222, 1991.

....Table 2 provides query length statistics for each of the topic sets and query representations. The goal of this research is to find a good strategy for combining query representations of varied length. Several approaches have been suggested for combining multiple query representations [2, 3] and more generally for fusing ranked sets, commonly called data fusion [4] These multiple evidence techniques (data fusion) are touted as a means to improve the effectiveness of information retrieval systems. They are based on the premise that repeated evidence increases the probability of ....

....along with the addition of term frequencies for cooccurring terms. Results of this strategy depend on the ranking algorithm to order the unified result set effectively. Turtle and Croft examined the effects of combining the evidence from two or more formal query requests in an inference engine [2]. Fox and Shaw [4] proposed several result combination algorithms and found that combinations of different ranking strategies yielded improvements in overall system effectiveness. One of their approaches was the CombSUM algorithm in which results from each strategy had their scores (min max) ....

H. Turtle, W. B. Croft, "Evaluation of an Inference Network-based Retrieval Model". ACM Transactions on Information Systems, 9:3, July 1991, pp. 187-222.

....to make the independence assumption more true, and 3) attempts to explain why the independence assumption isn t really needed anyway. Whatever its successes in machine learning, the first strategy has not met with great success in IR. While interesting research on dependence models has been done [8, 11, 21,49, 50], these models are rarely used in practice. Even most work in the inference net approach to information retrieval has mostly used independence (or ad hoc) models. Results from the second strategy are hard to judge. A variety of text representation strategies which tend to reduce independence ....

Howard R. Turtle and W. Bruce Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187-222, July 1991.

....different retrieval strategies) are likely to return different sets of relevant documents and different sets of non relevant documents. The development of this rationale is discussed in detail in [11] along with further discussion that, when taken in conjunction with Turtle and Croft s analysis [14] of multiple query representations, leads to a corollary of Belkin s original postulate. Namely, this corollary states that improvements might be seen from fusion even when the result sets are similar, as long as the difference in relevant overlap is greater than the difference in non relevant ....

H. Turtle and W.B. Croft, "Evaluation of an inference network-based retrieval model," ACM Transactions on Information Systems, Vol. 9, No. 3, pp. 187-222, 1991.

....properly estimated, it is well known that no other classifier can outperform Naive Bayes in the sense of misclassification probability. Attempts to overcome the restriction imposed by the independence assumption have motivated attempts to relax this assumption via a modification of the classifier [4], feature extraction in order to hold the assumption on stronger grounds, and approaches to underestimate the independence assumption by showing it doesn t make a big difference [3, 5] This paper is clearly on the second line of research: we propose a class conditional Independent Component ....

Turtle, H., Croft, W.: Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems 9 (1991) 187-222

....based of the rst class can be found in Section 3, here we will only address models of the second class. A probabilistic formalism for describing inference relations with uncertainty is provided by Bayesian inference networks, which have been described extensively in [68, 78] Turtle and Croft [98, 99] applied such networks to IR. Figure 1 depicts an example of such a network. Nodes represent IR entities such as documents, index terms, concepts, queries, and information needs. We can choose the number and kind of nodes we wish to use according to how complex we want the representation of the ....

H.R. Turtle and W.B. Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187-222, July 1991.

....number of retrieved sets in which the document appears. If the combining method is designed to favor the documents retrieved by more retrieval runs, the combined run can result in more accurate similarity values between queries and documents, and therefore give higher precision. Turtle and Croft [18] developed an inference network based retrieval model, which can combine different document representations and different versions of a query in a consistent probabilistic framework. Turtle and Croft implemented the INQUERY retrieval system based on the model, and demonstrated that multiple ....

H. Turtle and W.B. Croft, "Evaluation of an inference network-based retrieval model," ACM Transac- tions on Information Systems, Vol. 9, No. 3, pp. 187-222, 1991. 14

....prediction algorithm (WUFIS, Web User Flow by Information Scent) while firmly based in Information Foraging theory, is similar to several information retrieval algorithms based on network inferences. Turtle and Croft proposed the use of Bayesian networks to model information retrieval problems [25]. They represent queries and documents on an inference network, which is similar to our approach. More recently, a number of efforts in the Web research community have concentrated on combining linkage information with user queries in order to rank search results [6,5,7,14,24] Most similar to our ....

Turtle, H., Croft, W. (1991) Evaluation of an inference networkbased retrieval model. ACM Transactions on Information Systems, 9(3): 187-222

....does not allow for any form of relevance ranking of the retrieved document set, although some objects are more likely to be relevant or are more relevant to an information need than others. Excluding documents that do not precisely match a query specification results in lower effectiveness [Sal83, TC91] Therefore, the Boolean model is considered too weak for large text collections [TC92] Best match retrieval models have been proposed in response to the problems of exact match retrieval. These systems are based on the assumption that presenting documents to the user in order of presumed ....

....and in relevant and nonrelevant texts. The principle takes into account that the representation of both information need and text is uncertain, and the relevance relationship between them is also uncertain. Inference networks are one possible implementation of probabilistic models [TC90, TC91, BC92] Since inference networks can be used for implementing parts of the information filtering architecture that is the subject of this thesis (see Section 3.3.3) we will present this approach in some additional detail. Figure 3 shows the structure of an inference network consist ing of four ....

[Article contains additional citation context not shown here]

H.R. Turtle and W.B. Croft. Evaluation of an inference network-based retrieval model. ACM Trans. Inf. Syst., 9(3):187-222, July 1991.

....semantic indexing (LSI) probabilistic models (inference and belief networks) and connectionist networks. The vector space model [15] represents documents as weighted vectors and uses a cosine formula to evaluate similarity. Turtle and Croft have used inference networks to implement comparisons[17]. Newer models borrowing ideas from the AI field include Kwok s connectionist network[11] and Belew s AIR system[2] 3] Techniques such as LSI attempt to recognise latent interconnections occurring between terms[7] The effectiveness of the content based approach is constrained by inherent ....

H. Turtle and W. Croft. Evaluation of an inference network-based retrieval model. ACM Trans. on Info. Systems, 3, 1991.

....has proved beneficial in attempting to overcome the vocabulary problem, but has the disadvantage that the additional context provided by associated terms in a profile is ignored. 4. 5 Inference Networks The use of inference networks for information retrieval has been explored by Turtle and Croft [43]. A Bayesian inference network is an acyclic dependency graph in which nodes represent dependency relations between propositions. Given probabilities of the root nodes in this graph, probabilities for all remaining nodes may be calculated. The use of these networks for information retrieval ....

H. Turtle and W. Croft. Evaluation of an inference networkbased retrieval model. ACM Trans. on Info. Systems, 3, 1991.

....element being determined by some term weighting scheme. Retrieval involves ranking the document vector space with respect to the query (which has been located in the space as a result of indexing) based on some similarity function such as the cosine measure between vectors. Probabilistic methods [92, 111, 124] attempt to model the IR universe within a probabilistic framework based on the following assumption: given a query q and a document d in the collection, estimate the probability that the user will find the document d relevant. This estimation assumes that: 1) the probability of relevance depends ....

H. Turtle and W. B. Croft. Evaluation of an inference network-based retrieval model. ACM Trans. on Inf. Sys., 9(3):187, 1991.

....approaches in our study: the ranked list and interactive relevance feedback. The ranked list is the order of the retrieved documents as determined by INQUERY. The system uses an inference network model and estimates probabilities of how much each document satisfies user s information need [28]. The interactive relevance feedback procedure is as follows: we start from the top of the ranked list. Each time a new relevant document is discovered, we submit all the examined documents to the INQUERY s relevance feedback subsystem to modify the weights in the original query. Additionally, ....

H. Turtle and W. B. Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187--222, 1991.

....merged list of documents [4, 16, 17, 12] After a database selection algorithm ranked the 100 databases, the 10 most highly ranked databases were considered selected for search by that algorithm. The query used to rank databases was broadcast to Inquery IR systems serving each selected database [14, 3, 4], which each returned ranked lists of 100 documents. Each database was independent of the other databases; there was no exchange among databases of corpus statistics or other information, and no global corpus statistics were computed or maintained. The merging of document rankings produced from ....

H. R. Turtle and W. B. Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187--222, 1991.

....attached to the object network. This enables the computation of the probability that the information need is met for any particular object and, consequently, to produce a ranked list of objects. Figure II.5 presents a representation for such a scheme. More details of this process can be found in [Turtle and Croft, 1991]. o 1 o 2 o j 1 o . r 1 r 2 r 3 r m . q 1 q k . I Object Network Query Network Figure II.5. Basic inference network with the object and query networks. o j are object nodes, r m are concept representation nodes, q k are the query nodes and I represents the user s information need. ....

TURTLE, H. R. AND CROFT, W. B. 1991. Evaluation of an Inference Network Based Retrieval Model. In ACM Transactions on Information Systems, 9(3), pages187-222.

....parameters. Combining Sub Systems: The complexity of the information retrieval task has promoted the modularization of solutions, each potential solution specialized to deal with just a small part of the total complexity. For example, specialized approaches for dealing with phrases in documents [151], for handling synonyms [120] or for dealing with restricted kinds of queries [81] have been proposed. This specialization is also useful from a system designer s standpoint: Each specialized approach might be independently developed and maintained. However, this modularization also poses the ....

....retrieved 5 or more times was 6 times more likely to be relevant than a document retrieved in only one search. Additional research has demonstrated fairly convincingly that combining a number of different retrieval algorithms can result in performance improvements. For example, Turtle Croft [151] promote a method for combining the evidence from different experts using an inference network. Experts generate document relevance estimates which are propagated through the network and combined, using probabilistic inference rules, to form improved overall estimates. This model is used to ....

[Article contains additional citation context not shown here]

Howard Turtle and W. Bruce Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187--222, july 1991.

....0 or 1. The set of documents corresponding to units having an activation level near to 1 was retrieved for the input query. Performance of the inference network was evaluated using the CACM test collection. Retrieval effectiveness of the neural network was compared to that of a Bayesian network [58] based on precision and recall. The results showed that the neural network was able to give performance better (9.7 higher in average precision) than the Bayesian network. 32 3.4 Comparing the Different Approaches From the examples given in the last two sections, we can make the following ....

H. Turtle and W. B. Croft, "Evaluation of an inference network-based retrieval model," ACM Transactions on Information Systems, vol. 9, no. 3, pp. 187--222, 1991.

.... has the probability that document D is about concept T , The technique used by Turtle and Croft for assessing the belief in a representation concept is to assume that it it is proportional to the normalized within doccument frequercy tf , and to the collection frequency of the concept idf [38]. After testing several functions, they find that teh following equation is areasonable estimate on the CACM collection: p(T jD) 0:4 0:6 Delta tf Delta idf (22) A technique such as this one could easily be adapted to the PAS framework. Another problem is to find the probability that a ....

....contain more than n assumptions. This model of IR enters in the category of models where retrieval is seen as a form inference. There is a wide variety of approaches in that field, ranging from mostly symbolic ones to mostly numerical ones. One very popular approach is the inference network model [38], where relationships between document representation and query representation are modeled by a Bayesian network, and retrieval is done by computing the probability that the query concept is met given that some variables are observed. Advantages of PAS over IN are that PAS can deal with cycles. ....

[Article contains additional citation context not shown here]

H. Turtle and W.B. Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187--222, 1991.

....of query term occurrence patterns which are merged together using additive (vector space and probabilistic) or logical (boolean) models. The combination strategies seen in the literature range from simple unweighted boolean retrieval to sophisticated methods such as Bayesian inference networks [Turtle Croft, 1991] and logistic regression [Fuhr Pfeifer, 1994, Cooper et al. 1994] More recent interest has focused on the merging of higher level structures, such as multiple query or document representations (e.g. Bartell et al. 1994, Belkin et al. 1995] Perhaps the most common approach to method ....

.... representations (e.g. Bartell et al. 1994, Belkin et al. 1995] Perhaps the most common approach to method combination in information retrieval has been to test a series of carefully selected weighting functions [Belkin et al. 1995, Salton et al. 1983, Fox Shaw, 1994, Belkin et al. 1993, Turtle Croft, 1991, Schutze Pedersen, 1994, Lee, 1995] Many of these functions have parameters which are optimized by manual search over the parameter space using performance on IR test collections as the criterion to select the appropriate weights. A more principled approach is to use automatic methods to ....

Turtle, H., & Croft, W. B. (1991). Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3),187-- 222.

....the user is then initiated with the purpose of improving the probabilistic description of the ideal answer set through user feedback. 10 Beside these, alternative modeling paradigms such as fuzzy [106] extended Boolean retrieval models [79] latent semantic indexing [24] and neural networks [94, 100] have been proposed. Among these, the vector space model is one of the most popular and widely used in IR [7] In response to a request that information of a specific nature be selected, a search technique is generally employed, implemented by a Boolean or vector model. Documents are only ....

H. Turtle, and W. Croft, "Evaluation of an Inference Network-based Retrieval Model," ACM Transactions on Information Systems, vol. 9, no. 3, pp. 187-222, July 1991.

....to make the independence assumption more true, and 3) attempts to explain why the independence assumption isn t really needed anyway. Whatever its successes in machine learning, the first strategy has not met with great success in IR. While interesting research on dependence models has been done [8, 11, 21, 49, 50], these models are rarely used in practice. Even most work in the inference net approach to information retrieval has mostly used independence (or ad hoc) models. Results from the second strategy are hard to judge. A variety of text representation strategies which tend to reduce independence ....

Howard R. Turtle and W. Bruce Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187--222, July 1991.

No context found.

Howard Turtle and W. Bruce Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187--222, 1991.

No context found.

Turtle, H., and Croft, W. B. 1991. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems 9(3):187--222.

No context found.

H. Turtle and W. B. Croft. Evaluation of an inference networkbased retrieval model. TOIS, 9(3):187--222, 1991.

.... architecture of parameters and principles can be regarded as a realization of Universal Grammar (UG) Our experiments were implemented by a retrieval system called INQUERY (Callan, Croft Harding, 1992) which is based on the framework of a probabilistic inference network retrieval model (Turtle, 1991, and Turtle Croft, 1991) We already knew from previous studies that INQUERY is effective not only for English, but also for Japanese, Spanish, and Chinese. At this point, a questions arises: Where did INQUERY s cross linguistic applicability come from Obviously, the inference network is not a ....

.... of parameters and principles can be regarded as a realization of Universal Grammar (UG) Our experiments were implemented by a retrieval system called INQUERY (Callan, Croft Harding, 1992) which is based on the framework of a probabilistic inference network retrieval model (Turtle, 1991, and Turtle Croft, 1991). We already knew from previous studies that INQUERY is effective not only for English, but also for Japanese, Spanish, and Chinese. At this point, a questions arises: Where did INQUERY s cross linguistic applicability come from Obviously, the inference network is not a model based on ....

[Article contains additional citation context not shown here]

Turtle, H. & Croft, W. B. (1991). Evaluation of an inference network-based retrieval model, ACM Transactions on Information Systems, 9 (3), pp.

....problem solving effort. Since we are concerned with execution performance issues, the solution is fully implemented and evaluated empirically. The experimental test bed is provided by INQUERY [12] a full text probabilistic information retrieval system based on a Bayesian inference network model [88]. INQUERY was chosen for the following reasons: 9 INQUERY uses a general inverted file that includes term occurrence locations, allowing exploration of more complex inverted list data structures. This exploration would not be possible in a system that stores term weights only in its inverted ....

....in the network. Moreover, if our belief in any given proposition should change, its probability can be adjusted and the network can be used to update the probabilities at the rest of the nodes. The application of Baysien inference networks to information retrieval was advanced by Turtle and Croft [86, 88, 87]. The inference network used for information retrieval is 98 I d 1 d 2 d n t 1 t 2 t k r 1 r 2 r 3 r v c 1 c 2 c 3 c j q 1 document network query network q 2 Figure 4.1 Inference network for information retrieval divided into two parts, a document network and a query network, ....

[Article contains additional citation context not shown here]

Turtle, H. R. and Croft, W. B. Evaluation of an inference network-based retrieval model. ACM Trans. Inf. Syst., 9(3):187--222, July 1991.

No context found.

H. Turtle and W. B. Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187--222. ACM, July 1991.

No context found.

H. Turtle and W. B. Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187--222, 1991.

No context found.

Turtle, H., Croft, W.B.: Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems (TOIS) 9 (1991) 187--222

No context found.

H. Turtle and W. B. Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187--222, 1991.

No context found.

H. Turtle and W. B. Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187--222. ACM, July 1991.

No context found.

Turtle, H. and Croft, B., "Evaluation of an Inference Network-Based Retrieval Model," ACM Transaction on Information Systems , Vol. 9, no. 3, July 1991, pp. 187-222.

No context found.

H. Turtle and W. B. Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187--222, 1991.

No context found.

Turtle, H. and Croft, W. B. (1991). Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187--222.

No context found.

Turtle, H. and Croft, W. B. (1991). Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187--222.

No context found.

Turtle, H. and Croft, W. B. (1991). Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187--222.

No context found.

H. R. Turtle and W. B. Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9:187--222, 1991. 421

No context found.

H. Turtle and W. B. Croft. Evaluation of an inference network-based retrieval model. ACM TOIS, 9:187--222, 1991.

No context found.

H. R. Turtle and W. B. Croft. Evaluation of an Inference Network-Based Retrieval Model. ACM Transactions On Information Systems, 9(3):187--222, 1991.

No context found.

Turtle, H. and Croft, W. (1991). Evaluation of inference network-based retrieval methods. ACM Transactions on Information Systems, 9(3):187--222.

No context found.

Turtle, H., Croft, W. (1991) Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187-222

No context found.

H. R. Turtle and W. B. Croft. Evaluation of an Inference Network-Based Retrieval Model. ACM Transactions On Information Systems, 9(3):187--222, 1991.

No context found.

Howard R. Turtle and W. Bruce Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187--222, 1991.

No context found.

H. Turtle and W.B. Croft. "Evaluation of an inference network-based retrieval model." ACM Transactions on Information Systems, 9,3: 187-222, 1991.

No context found.

H. Turtle and W. B. Croft, Evaluation of an inference network-based retrieval model, ACM Transactions on Information Systems, 1995, 9: 187-222.

No context found.

Turtle, H. & Croft, W.B. (1991). Evaluation of an inference network-based retrieval model.

No context found.

H. R. Turtle and W. B. Croft. Evaluation of an inference network-based retrieval model. ACM Transactions on Information Systems, 9(3):187-222, 1991.

No context found.

Turtle, H. and W. B. Croft, "Evaluation of an Inference Network based Retrieval Model", ACM Transactions on Information Systems , Vol. 9, No. 3, July 1991.

First 50 documents  Next 50

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC