C. Buckley and G. Salton. Optimization of relevance feedback weights. In E. Fox, P. Ingwersen, and R. Fidel, editors, Proceedings of the 18th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 351-- 357, 1995.  Home/Search   Document Not in Database   Summary   Related Articles   Check

.... schemes for CBIR have been mainly adopted from text document retrieval research and can be classified into two approaches: query point movement (query refinement) and re weighting (similarity measure refinement) Both have been built based upon the vector model in information retrieval theory [13, 14]. Recently, more computationally robust methods that perform global optimization have been proposed. The MindReader retrieval system formulates a minimization problem on the parameter estimating process [9] It allows for correlations between attributes in addition to different weights on each ....

C. Buckley, G. Salton, Optimization of Relevance Feedback Weights, Proc. 18 Annual Intl ACM SIGIR Conf., Seattle, USA, 1995, pp. 351-357.

....describe user profiles and the way the profiles are maintained. Each term in a user profile expresses, in a certain degree, the likes of a particular user. A weight is associated to each term for denoting this degree. This weight indicates the importance of the term in the users interests [1,15]. The weights dynamically change as a result of the users behavior, and so the importance of that term in the users interests changes as well [1,15,16,17,18,19,20,21,22] Moreover, if the weight of a term becomes zero, that term is removed from the profile. According to their associated weights, ....

....particular user. A weight is associated to each term for denoting this degree. This weight indicates the importance of the term in the users interests [1,15] The weights dynamically change as a result of the users behavior, and so the importance of that term in the users interests changes as well [1,15,16,17,18,19,20,21,22]. Moreover, if the weight of a term becomes zero, that term is removed from the profile. According to their associated weights, all terms in a profile are classified into two groups: one that contains heavier terms (called LONG TERMS) and one containing the lighter ones (called SHORT TERMS) ....

# C. Buckley and G. Salton. Optimization of Relevance Feedback Weights. In Proceedings of the #8th Annual Intl ACM SIGIR Conference, Seattle, 1995.

....in experimental systems explicit feedback has the added advantage of minimizing one potential source of experimental error, inference of the user s true reaction. Several papers exhibit the outperform of the systems achieved with the explicit relevance feedback ( Salton and Buckley, 1990] and [Buckley and Salton, 1995]) Figure 10. Explicit Relevance Feedback But in practical applications explicit feedback has three serious drawbacks: First, the relevance of information is always relative to the changing information need of a user, and information environments relevance judgements of individual items are ....

Buckley, C., Salton, G., "Optimization of relevance feedback weights". In Fox, Ed, Ingwersen, Peter, and Fidel, Raya (Editors), Proceedings of the Eighteenth Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 351--357. 1995.

....a query object (associated with it an attribute vector) database objects whose attribute vectors are most similar to the query vector are retrieved for the user. Usually k top matches are retrieved. For text applications, vector space model with cosine similarity metric is being widely used [17] [4]. The cosine similarity is defined as, S(#x, #q) #x.#q #x #q , where #u.#v stands for inner product of #u and #v, and #v denotes the magnitude of #v. On the other hand, in metric space model of information retrieval, secondorder (L 2 ) distance metrics are typically used. The ....

C. Buckley, and G. Salton. Optimization of relevance feedback weights. In Proc. of SIGIR, pages 351--357, Seattle, WA, 1995.

....in the mid 1960 s to improve queries using relevance feedback. It has proven to be one of the most successful feedback algorithms. Rocchio [28] showed that the optimal query vector is the di erence vector of the centroid vectors for the relevant and the non relevant documents. Salton and Buckley [4] included the original query (Q orig ) to preserve the focus of the query, and added coecients ( and ) to control the contribution of each component. The mathematical formulation of this version is: Q new = Q orig R d2rel d N R d=2rel d (12) where d is the weighted ....

....documents, and N is the total number of documents. Any negative components of the nal vector Q new are set to zero. Several techniques have been proposed to improve the e ectiveness of Rocchio s method: better weighting schemes [33] query zoning [34] and dynamic feedback optimization [4]. As pointed out by Schapire et al. 31] most of the studies that use Rocchio as a baseline have constructed a weak version of the classi er [16, 19, 42, 44] They also show that a properly optimized Rocchio s algorithm could achieve quite competitive performance. We have noticed that Rocchio s ....

Buckley C and Salton G. Optimization of relevance feedback weights. In Proceedings of the 18th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 351-357, Seattle, WA, July 1995.

....to the feedback examples. Generally speaking, previous RF methods can be classified into two categories: re weighing approach and probability approach . Most of existing works in CBIR use the former approach, notably based on a popular model in information retrieval called the vector model [1][17] 18] which is adopted in some systems [11] 14] 16] 7] 10] The idea is to associate larger weights with more important dimensions and small weights with unimportant ones. This can be illustrated by the Rocchio s formula. For a set of relevant documents and nonrelevant documents given by the ....

.... vector is [ in x x x , 1 = r ,where ) 3 ) m m m m x x x x i = and ) min( max( min( m m m m m i X X X x x = If m i x satisfies the Gaussian distribution, it is easy to prove that the probability of m i x being in the range of [ 1,1] is 99 . 2. Initialization: Initialize I = i (identity matrix) and i = x , n =1. Retrieval and Feedback 1. Update the retrieval parameters i , i ,andn according to Equation (8) 10) using the information provided by the current set of positive examples U. 2. Distance Calculation: For ....

Buckley, C., and Salton, G. "Optimization of Relevance Feedback Weights," in Proc of SIGIR'95.

.... ] 1 n i k k i k x x x = where ) 3 ) m m m m k k k k x x x x = and ) min( max( min( m m m m m k i k k i k i k X X X x x = If m i k x satisfies the Gaussian distribution, it is easy to prove that the probability of m i k x being in the range of [ 1,1] is 99 . 2. Initialization: Initialize i k to be null and let i k = i k x , k n =1. 3. Feedback and Update Parameters according to equation (5) 6) and (7) 4. Distance Calculation: For each image i P in the database, we calculate k i d , its distance away from the example image k P , by ....

Buckley, C., and Salton, G. "Optimization of Relevance Feedback Weights," in Proc of SIGIR'95.

....algorithm also has the progressive learning capability that utilize past feedback information to help the current query. Experimental results show that our algorithm is effectiveness. 1. Introduction While there are many research efforts in addressing content based imaged retrieval (CBIR) [1][5] 8] the performance of CBIR methods are still limited. One of the problems that pose performance limitations to CBIR is the disparity between semantic concepts and low level image features. The mapping between them is still impractical with today computer vision and AI techniques. To improve ....

....query point by moving it towards positive examples and away from negative ones. The frequently used technique to iteratively improve this estimation is the Rocchio s formula [7] Experiments show that the retrieval performance can be improved considerably with such relevance feedback approaches [1][9] 10] The key idea behind the re weighting method is very simple and intuitive. The MARS[7] system implemented a refinement to re weighting methods, called standard deviation method. Each image can be viewed as a vector in a N dimensional feature space; and if the variance of positive examples ....

[Article contains additional citation context not shown here]

Buckley, C., and Salton, G. "Optimization of Relevance Feedback Weights," in Proc of SIGIR'95.

....sets the weight of a query feature or term according to the product of term frequency tf and inverse document frequency idf. Robertson and Sparck Jones [17] derived a variant of the tf idf formula using a probabilistic theory of relevance weighting. The third method uses Rocchio weighting formula [5, 18] which modi es a term weight in the query in proportion to the positive sum of the corresponding term weights in relevant documents and the negative sum of term weights in irrelevant documents. The above methods tend to be somewhat ad hoc or provide only partial solutions [20] Rui and Huang [20] ....

C. Buckley and G. Salton. Optimization of relevance feedback weights. In Proc. SIGIR '95, pages 351-357, 1995.

....the experimental results are given in section 5. Finally, section 6 is the summery. 2. Relevance Feedback Relevance feedback is a technique that takes advantage human computer interaction to refine high level queries represented by low level features. It is used in traditional document retrieval [1] for automatically adjusting an existing query using information fed back from the user. In the application of image retrieval [6] the user selects relevant images from previous retrieved results and provides a preference weight for each relevant image. The weights for the low level feature, ....

C. Buckley and G. Salton, "Optimization of relevance feedback weights", in Proc. Of SIGIR'95.

....such as SQL, that treats both scenarios in a uniform way and provides support for approximate matching. 3.1 Syntax The fuzzy SQL language is conjunctive SQL (or conjunctive relational calculus (see [34] extended to accommodate approximate matching. The SQL syntax is extended in two ways ([2, 35]) The first extension generalizes the definition of atoms (or atomic conditions) 34] to include the following approximate comparison operators: 2 f = g, i.e. approximate equal , approximatebigger , approximate smaller . An atom is called an approximate atom if it contains an ....

Buckley C., Salton G., Optimization of Relevance Feedback Weights, in Proceedings of the 18th Annual Intl ACM SIGIR Conference, Seatle, 1995.

....Schemes Used in Information Retrieval 13 Of course, variations on the previously described relevance feedback scheme have been suggested. For example, Robertson (1990) suggests adopting different procedures for the selection of new search terms and for the weighting of search keywords, see also (Buckley Salton 1995). For Harman (1992) the number of terms to be included in the expanded request can depend on the average document size of the underlying collection, and Allen (1995) seems to confirm this finding. For Aalbergsberg (1992) the number of documents shown to the user must not be an arbitrary number ....

Buckley, C. & Salton, G. (1995). Optimization of relevance feedback weights.

....results and conclusions are given in Section 4 and 5, respectively. 2. Relevance Feedback in CBIR 2. 1 Relevance Feedback Relevance feedback is a process of automatically adjusting an existing query using information fed back from the user about the relevance of previous retrieved document [6]. In MARS [2] the user tells the system about which images of the previous retrieved results are relevant to what he or she wants and provides a preference weight for each relevant image. Query weights for each low level feature, i.e. color and texture, etc. are dynamically updated based on the ....

C. Buckley and G. Salton, "Optimization of relevance feedback weights", in Proc. Of SIGIR'95

....attribute value (such as young ) was introduced in [30] The modification is motivated by the fact that each fuzzy term may have a different meaning in the query. However, this modification does not involve relevance feedback. Relevance feedback methods have been extensively studied in IR area ([2, 3, 22]) For the text part of our query refinement approach, we assume that we can plug in any IR system that supports relevance feedback, such as SMART [25] Most of the research using relevance feedback in a multimedia retrieval system is related to the MARS system (see [19, 20, 23] and MindReader ....

Buckley C., Salton G., "Optimization of Relevance Feedback Weights", in Proceedings of the 18th Annual Intl ACM SIGIR Conference, Seatle, 1995.

....use distribution statistics (IDF) from a reference corpus instead of the target test corpus. The ranks and scores determined by term selection methods are expected to reflect the relative importance of terms in the profile. However, as already observed and explored by other TREC participants (see [Buckley Salton 1995]) and as we experienced in our preliminary experiments, optimal profile terminology weighting does not follow directly from the term selection process. Furthermore, it remains a challenge to incorporate term weights appropriately into the relevance matching function in vector space retrieval. The ....

Buckley, Chris, and Salton, Gerard, "Optimization of relevance feedback weights". In Fox, Ed, Ingwersen, Peter, and Fidel, Raya (Editors), Proceedings of the Eighteenth Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 1995, 351--357.

....set of labeled examples may be used to bias a clustering algorithm; we investigate how a user, interacting with the system, may eciently guide the learner to a desired clustering. Our technique of incorporating user feedback is a cousin to relevance feedback, a technique for information retrieval (Buckley Salton, 1995). Given a query and initial set of retrieved documents, relevance feedback asks the user to tag documents as being more or less relevant to the query being pursued. As the process is iterated, the retrieval system builds an increasingly accurate model of what the user is searching for. The ....

Buckley, C. & Salton, G. (1995) Optimization of Relevance Feedback Weights. SIGIR 1995.

....to build true Image Databases [101, 44] the systems are not at the complete stage yet. A successful Image Database System requires an interdisciplinary research e ort. Besides the integration of Database Management and Computer Vision, the research from the traditional Information Retrieval area [16, 126, 132, 125, 15, 7] is also an indispensable part. Although the traditional Information Retrieval area s research focus was in text based document retrieval, many useful retrieval models and techniques can be adapted to Image Retrieval. Some successful examples of such research e ort include the adaption of Boolean ....

Chris Buckley and Gerard Salton. Optimization of relevance feedback weights. In Proc. of SIGIR'95, 1995.

....this claim, we show that our approach signi cantly outperforms baseline IR approaches that could be applied to the same task. We compare the performance of language modeling to a more traditional vectorspace (cosine similarity) approach commonly used in Information Retrieval for text classi cation [2]. Brie y, a vector space approach compares each document D in the test set to the centroid C t of the training documents associated with trend t. Both C t and D are viewed as vectors in vocabulary space, and cosine of the angle between the two vectors constitutes similarity. We used Okapi tf idf ....

Chris Buckley and Gerard Salton. Optimization of relevance feedback weights. In Proceedings on the 18th annual international ACM SIGIR conference, pages 351-357, 1995.

....method and provide experimental evaluations showing its effectiveness in image retrieval. Concluding remarks will be given in Section 5. 2. RELATED WORK This work was performed at Microsoft Research China. One of the most popular models used in information retrieval is the vector model [1, 8, 9]. Various effective retrieval techniques have been developed for this model and among them is the method of relevance feedback. Most of the previous relevance feedback research can be classified into two approaches: query point movement and re weighting [3] The query point movement method ....

....D N D N Q Q g b a where a, b, and g are suitable constants; N R and N N are the number of documents in D R and D N respectively. This technique is implemented in the MARS system [6] Experiments show that the retrieval performance can be improved considerably by using relevance feedback [1, 8, 9]. The central idea behind the re weighting method is very simple and intuitive. The MARS system mentioned above implements a slight refinement to the re weighting method call the standard deviation method [6] Since each image is represented by an N dimensional feature vector, we can view it as a ....

Buckley, C., and Salton, G. "Optimization of Relevance Feedback Weights," in Proc of SIGIR'95.

....[12, 11] embody this view, emphasizing the construction of a prototypical relevant vector to which similarity can be measured. In recent years, IR researchers working in both frameworks have used increased computing power to search for models that optimize ranking effectiveness on training data [1, 3, 10]. The resulting algorithms have much in common with techniques from machine learning but lack the theoretical analyses (such as proofs of convergence) often pursued in machine learning. Conversely, machine learning has devoted relatively little attention to ranking, although some areas of ....

....are multipass optimization algorithms initialized using Rocchio s relevance feedback formula [12, 11] We therefore compared the effectiveness of RankBoost with the latest in this series of algorithms. This version, which we will call Rocchio QZ DFO here, incorporates dynamic feedback optimization [3] and query zoning [18] The algorithm has 5 parameterized phases and is described in detail elsewhere [16] 6. EXPERIMENTS We now report our experimental results on applying RankBoost to train ranking models. Study 1 focuses on issues of model fitting, while Study 2 focuses on the character of ....

Chris Buckley and Gerard Salton. Optimization of relevance feedback weights. In Proceedings of the 18th Annual International Conference on Research and Development in Information Retrieval, pages 351-- 357, July 1995.

....of little value for retrieval. Commonly the inverse of the logarithm of df is taken (log( N doc df ) where N doc is the number of documents in the collection) and named inverse document frequency (idf) All terms i in a document are then assigned weights according to: w i = tf i Theta idf i [3, 24, 26], words not present have a weight of 0. Let Nword be the number of distinct words in the document collection, each document is then viewed as a point (vector) in an Nword dimensional space. 3 Queries and documents are represented the same way, weights are extracted for all the terms in a query ....

....however the relevance feedback obtained from the user provides reasonable approximations to these sets. Information fed back is used to construct a new query vector which is closer to the optimum query. Experiments show that the retrieval performance improves considerably with this technique [3, 24, 26]. The technique described moves only the center point of the query space to a new centroid. Figure 4a) shows a two dimensional example, where the query point ( Theta) is moved to a location central to the three relevant vectors. The query vector changed, but not the similarity function. By ....

C. Buckley and G. Salton. Optimization of relevance feedback weights. In Proc. of SIGIR '95, 1995.

....outperforms the fuzzy model in terms of retrieval performance (discussed in section 4) 3.4 Vector Model An IR model consists of a document model, a query model, and a model for computing similarity between the documents and the queries. One of the most popular IR models is the vector model [Buckley and Salton, 1995, Salton and McGill, 1983, Shaw, 1995] Various effective retrieval techniques have been developed for this model. Among them, term weighting and relevance feedback are of fundamental importance. 3.4.1 Term Weighting in Textual Media Term weighting is a technique for assigning different weights ....

....inversely with the number of documents in which a term appears. idf k = log 2 M df k 1 (11) where df k is the document frequency for term t k and M is the total number of documents in the collection. Experiments have shown that the product of tf and idf is a good estimation of the weights [Buckley and Salton, 1995, Salton and McGill, 1983, Shaw, 1995] The query Q has the same model as that of a document D, i.e. it is a weight vector in the term space: Q = w q1 ; w qk ; w qN ] 12) The similarity between D and Q is defined as the Cosine distance. Similarity(D; Q) D Theta Q jjDjj Theta ....

[Article contains additional citation context not shown here]

Buckley, C. and Salton, G. (1995). Optimization of relevance feedback weights. In Proc. of SIGIR'95.

....of the term. To denote the penalty for a high document frequency, this figure is usually inverted and named inverse document frequency (idf ) and is computed as: idf = log collection size df (2. 2) Experiments have shown that the product of tf and idf is a good estimation of the weights [6, 38, 40]. The final weight w i for term i in the document is: w i = tf i Theta idf i (2.3) 8 Now, each document has a weight assigned for all possible words (words not in the document have a weight of 0) If there are Nword distinct words in the collection of all documents, then each document can now ....

....back from the user, the system can adjust the position of the point in the vector space to closer approach the ideal position. If the set of relevant feature vectors (F relevant ) and non relevant feature vectors (F non Gammarelevant ) are known, the optimal query vector V opt can be proven to be [6, 38, 40]: V opt = 1 N relevant X i2F relevant F i Gamma 1 N collection Gamma N relevant X i2F non Gammarelevant F i (2.27) where N relevant is the number of vectors in F relevant and N collection the number of the total feature vectors in the collection. In practice, F relevant and F ....

[Article contains additional citation context not shown here]

Chris Buckley and Gerard Salton. Optimization of relevance feedback weights. In Proc. of SIGIR'95, 1995.

....an existing query using the information fed back by the user about the relevance of previously retrieved objects. By incorporating relevance feedback into the retrieval process, human perception subjectivity can be better modeled, thus resulting in considerable improvement in retrieval performance [1,2,3]. In MIR, the issue of human perception subjectivity is even more important than that in TIR because of the rich multimedia content contained in the multimedia objects. Development of techniques that can incorporate human perception subjectivity into MIR is thus crucial for a successful MIR ....

Buckley and G. Salton, "Optimization of relevance feedback weights", in Proc. Of SIGIR'95

....w i (Q # ) are the weights of the same term in the old and new queries. Parameters #, #, and # called Rocchio coe#cients control the relative impact of each component. Generally, these parameters are selected empirically and the best 1020 terms are added to the query [4] Buckley and Salton [9] suggested an approach called Dynamic Feedback Optimization (DFO) where the coe#cients are learned by greedy exploration of the parameter space. They also demonstrated that increasing the number of expansion terms can improve performance. Biron and Kraft [8] give a detailed overview of ....

....process progresses and achieves a 9 improvement. The third search strategy that uses tiling to represent its function demonstrates a 10 improvement. Table 4: Rocchio coe#cients from three di#erent sources. System Old query (# 1 ) Relevant (# 2 ) Non rel. # 3 ) Inquery [3] 0. 5 4 1 DFO [9] 0.25 8 1 F 1 (D, d) 0.5 2 1 We observed a similar increase in average precision while considering the top 100 documents instead of the top 50 (Table 2) The search strategies were beginning their exploration from a known starting point: the top ranked relevant document and all non relevant ....

[Article contains additional citation context not shown here]

C. Buckley and G. Salton. Optimization of relevance feedback weights. In Proceedings of ACM SIGIR, pages 351--357, 1995.

....results suggest that query specific tailoring of these numbers may not be effective. 16, 17] 4. 2 Weight Assignment A step popular in relevance feedback methodology is to assign query term weights based on a closedform function originally developed by Rocchio[9] and has been improved upon in [10, 7, 11]. The 4 weight assigned to a concept added to an expanded query is 8 tf rel Gamma 2 tf nonrel , where tf rel is the average tf component of the concept in relevant documents, and tf nonrel is the average tf component in non relevant documents. Recent research[7, 11, 2] indicates that iterative ....

....been improved upon in [10, 7, 11] The 4 weight assigned to a concept added to an expanded query is 8 tf rel Gamma 2 tf nonrel , where tf rel is the average tf component of the concept in relevant documents, and tf nonrel is the average tf component in non relevant documents. Recent research[7, 11, 2] indicates that iterative techniques can be used to improve these weights. 5 Experiments Experiments were conducted on 50 natural language information requests used for the routing track for TREC 4[1] The information requests were stopped and stemmed to produce an initial query, and ....

[Article contains additional citation context not shown here]

C. Buckley and G. Salton, "Optimization of Relevance Feedback Weights", Proceedings of SIGIR 1995, pp. 351-357.

....the query and threshold over time. We also tested several weight learning approaches as extensions to the static query formulation process. Our results suggested that weight learning approaches such as Widrow Hoff [10] Exponentiated Gradient Descent [5] and Dynamic Feedback Optimization [1] are of limited use for event tracking. It was observed that the 1 http: www.ldc.upenn.edu Table 1: NIST evaluation of TDT2 Tracking Systems (Story Weighted Cost, Nt=4) TDT2 Cost P (m) P (fa) UPENN 0.0058 0.0934 0.0040 UMASS A 0.0059 0.0855 0.0043 BBN 0.0063 0.1415 0.0035 UMASS S 0.0070 ....

C. Buckley and G. Salton, "Optimization of Relevance Feedback Weights," Proceedings of ACM SIGIR, 351-357, 1995.

....Bartell [7] adds significant evidence that suggests this hypothesis is true. Earlier contributions include a widely used query expansion technique introduced by Rocchio [70] This method has been improved using machine learning approaches [48, 49, 56, 80, 78] and heuristic optimization techniques [10] on very large collections. Most of the previous work for document classification has focused on supervised methods that use training documents with known relevance. Since relevance assessments are not always available, it is desirable to improve retrieval effectiveness without them. A paradigm ....

....in more detail in Section 4.1.2.1. We also experimented with feature selection and weight assignment variants of the baseline process. We tested static classifiers expanded with multiword features (MWF) 57] We also tested two weight learning algorithms: Dynamic Feedback Optimization (DFO) [10] and Exponentiated Gradient Descent (EG) 41] In addition to comparing extensions to static classifier formulation, one of the goals of our tracking experiments is to determine the impact that automatic speech recognition (ASR) technology has on our process. We compare results using ASR data to ....

[Article contains additional citation context not shown here]

C. Buckley and G. Salton, "Optimization of Relevance Feedback Weights," Proceedings of ACM SIGIR, 351-357, 1995.

....to build true Image Databases [102, 45] the systems are not at the complete stage yet. A successful Image Database System requires an interdisciplinary research effort. Besides the integration of Database Management and Computer Vision, the research from the traditional Information Retrieval area [17, 127, 133, 126, 16, 8] is also an indispensable part. Although the traditional Information Retrieval area s research focus was in text based document retrieval, many useful retrieval models and techniques can be adapted to Image Retrieval. Some successful examples of such research effort include the adaption of Boolean ....

Chris Buckley and Gerard Salton. Optimization of relevance feedback weights. In Proc. of SIGIR'95, 1995.

....and conclusions will be given in sections 4 and 5 respectively. 2. TERM WEIGHTING AND RELEVANCE FEEDBACK An IR model consists of a document model, a query model, and a model for computing similarity between the documents and the queries. One of the most popular IR models is the vector model[14, 3, 15]. Various effective retrieval techniques have been developed for this model. Among them, term weighting and relevance feedback are of fundamental importance. 2.1. Term Weighting Term weighting is a technique of assigning different weights for different keywords (terms) according to their ....

....varies inversely with the number of documents in which a term appears. idf k = log2 M dfk 1 (2) where dfk is the document frequency for term k and M is the total number of documents in the collection. Experiments have shown that the product of tf and idf is a good estimation of the weights[14, 3, 15]. The query Q has the same model as that of document D, i.e. it is a weight vector in the term space: Q = wq1 ; w qk ; wqN ] 3) The similarity between D and Q is defined as the Cosine distance. Sim(D; Q) D Q jjDjj jjQjj (4) where jj jj denotes norm 2. 2.2. Relevance Feedback ....

[Article contains additional citation context not shown here]

C. Buckley and G. Salton, "Optimization of relevance feedback weights," in Proc. of SIGIR'95.

....a correlation between increasing cooccurrence ratios and improved precision. Our analysis has addressed the impact of feature cooccurrence, but we have yet to address the effects of feature weight assignment. We anticipate that feature weight learning approaches applied to single word features [4, 14, 15] will further improve retrieval effectiveness for queries with multiword features. We are continuing to investigate the value of phrases and multiword features for several document classification tasks. Our belief is that query expansion approaches are limited when only independent single word ....

C. Buckley and G. Salton, "Optimization of Relevance Feedback Weights," Proceedings of SIGIR, 351-357, 1995.

No context found.

C. Buckley and G. Salton. Optimization of relevance feedback weights. In E. Fox, P. Ingwersen, and R. Fidel, editors, Proceedings of the 18th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 351-- 357, 1995.

No context found.

Chris Buckley and Gerard Salton. Optimization of relevance feedback weights. In SIGIR, pages 351--357, 1995.

No context found.

Buckley, C. and Salton, G.: 1995, Optimization of relevance feedback weights, Proceedings of 18th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 351357.

No context found.

C. Buckley and G.Salton, "Optimization of Relevance Feedback Weights," in Proc of SIGIR 1995.

No context found.

C. Buckley and G.Salton, "Optimization of Relevance Feedback Weights," in Proc of SIGIR 1995.

No context found.

C. Buckley and G. Salton, "Optimization of relevance feedback weights," in Proc. SIGIR'95.

No context found.

C. Buckley and G. Salton. Optimization of relevance feedback weights. In Proceedings of the 18th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 351--357, New York, NY, July 1995. ACM press.

No context found.

C. Buckley and G. Salton. Optimization of relevance feedback weights. In Proceedings of SIGIR-95, pages 351--357, Seattle, US, 1995.

No context found.

C. Buckley and G. Salton, "Optimization of relevance feedback weights," in Proc. SIGIR'95.

No context found.

Buckley C., Salton G. "Optimization of Relevance Feedback Weights". In Proceedings of the 18th Annual Intl ACM SIGIR Conference, Seatle, 1995.

No context found.

C. Buckley and G. Salton. Optimization of relevance feedback weights. In Proceedings of SIGIR-95, pages 351--357, Seattle, US, 1995.

No context found.

Buckley, C., and Salton, G. Optimization of relevance feedback weights. In Proceedings of the 18th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (Seattle, July 1995).

No context found.

C. Buckley and G. Salton, "Optimization of relevance feedback weights," presented at the SIGIR'95.

No context found.

C. Buckley and G. Salton. Optimization of relevance feedback weights. In Proc. of SIGIR95, the 18th ACM Conference on Research and Development in Information Retrieval, pages 351357, Seattle, USA, July 1995.

No context found.

C. Buckley and G. Salton. Optimization of relevance feedback weights. In Proceedings of the Eighteenth Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 351--357. Fox, Ed, Ingwersen, Peter, and Fidel, Raya (Editors), 1995.

No context found.

C. Buckley and G. Salton, "Optimization of relevance feedback weights," in Proc. of SIGIR'95.

No context found.

Chris Buckley and Gerard Salton. 1995. Optimization of relevance feedback weights. In Proceedings of the 18th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 351-357.

No context found.

C. Buckley and G. Salton, #Optimization of relevance feedbackweights," in Proc. of SIGIR'95, 1995.

No context found.

C. Buckley and G. Salton, "Optimization of relevance feedback weights," in Proc. of SIGIR'95, 1995.

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