Ilmério Silva

Link-based and content-based evidential information in a belief network model

This work presents an information retrieval model developed to deal with hyperlinked environments. The model is based on belief networks and provides a framework for combining information extracted from the content of the documents with information derived from cross-references among the documents. The information extracted from the content of the documents is based on statistics regarding the keywords in the collection and is one of the basis for traditional information retrieval (IR) ranking algorithms. The information derived from cross-references among the documents is based on link references in a hyperlinked environment and has received increased attention lately due to the success of the Web. We discuss a set of strategies for combining these two types of sources of evidential information and experiment with them using a reference collection extracted from the Web. The results show that this type of combination can improve the retrieval performance without requiring any extra information from the users at query time. In our experiments, the improvements reach up to 59% in terms of average precision figures.

Local versus global link information in the Web

Information derived from the cross-references among the documentsin a hyperlinked environment, usually referred to as linkinformation, is considered important since it can be used toeffectively improve document retrieval. Depending on the retrievalstrategy, link information can be local or global. Local linkinformation is derived from the set of documents returned asanswers to the current user query. Global link information isderived from all the documents in the collection. In this work, weinvestigate how the use of local link information compares to theuse of global link information. For the comparison, we run a seriesof experiments using a large document collection extracted from theWeb. For our reference collection, the results indicate that theuse of local link information improves precision by 74%.When global link information is used, precision improves by35%. However, when only the first 10 documents in theranking are considered, the average gain in precision obtained withthe use of global link information is higher than the gain obtainedwith the use of local link information. This is an interestingresult since it provides insight and justification for the use ofglobal link information in major Web search engines, where usersare mostly interested in the first 10 answers. Furthermore, globalinformation can be computed in the background, which allowsspeeding up query processing.

Bayesian belief networks for IR

We review the application of Bayesian belief networks to several information retrieval problems, showing that they provide an effective and flexible framework for modeling distinct sources of evidence in support of a ranking. To illustrate, we explain how Bayesian networks can be used to represent the classic vector space model and demonstrate how this basic representation can be extended to naturally incorporate new evidence from distinct information sources. These models have been shown useful in several text collections, where the combination of evidential information derived from past queries, thesauri, and the link structure of Web pages has led to significant improvements in retrieval performance.

Bayesian Network Models for Information Retrieval

In this chapter, we apply Bayesian networks to the problem of retrieving information about a subject or topic and show that Bayesian networks provide an effective and flexible framework for dealing with information retrieval (IR) in general. Our discussion focus on two Bayesian networks models proposed in the literature namely, the inference network and the belief network models. We compare the expressiveness of these two models and show that the belief network model is more general. We also demonstrate that the belief network model is general enough to subsume the three classic IR models namely, the Boolean, the vector, and the probabilistic models. Further, we show that a belief network can be used to naturally incorporate pieces of evidence from past user sessions which leads to improved retrieval Performance. At the end, for comparative purposes, we review models of reasoning other than the Bayesian networks and characterize a taxonomy for them.

Spatial Locality Weighting of Features Using Saliency Map with a Bag-of-Visual-Words Approach

In this paper we propose a new descriptor for content-based image retrieval that explores the locality of features. We propose to extend the bag-of-visual-words method by weighting the visual words according to their spatial locality in terms of foreground and background by using fuzzy saliency models. We evaluated our method using databases that obtains images with different conditions of illumination, color, rigid and scale transformations, and changes of the background. The analysis of the results demonstrated that our proposal presents significant improvements over competitive approaches.

Combining Color and Topology for Partial Matching

Although, color is one of the most visually distinguishable visual properties, color alone is not enough to describe the content of images. The spatial organization of the different color regions also play an important role. In this paper, we propose and evaluate a new descriptor that combines information about color and about its spatial arrangement in an image. Moreover, the mechanism used to compute the descriptor provides support for partial matching of images and for the development of efficient retrieval systems. We first describe the spatial arrangement of the color regions using a topological graph, where vertices represent the color regions and edges represent connections between regions and also the color differences between them. To compute the descriptor from this graph representation we use the spectral graph theory, avoiding the need for direct graph comparison. We performed various experimental evaluations to compare the accuracy of our new descriptor with descriptors based only on color, based only on topological information and a combination of both.

Comparison of representations of multiple evidence using a functional framework for IR

The combination of sources of evidence is an important subject of research in information retrieval and can be a good strategy for improving the quality of rankings. Another active research topic is modeling and is one of the central tasks in the development of information retrieval systems. In this paper, we analyze the combination of multiple evidence using a functional framework, presenting two case studies of the use of the framework to combine multiple evidence in contexts bayesian belief networks and in the vector space model. This framework is a meta-theory that represents IR models in a unique common language, allowing the representation, formulation and comparison of these models without the need to carry out experiments. We show that the combination of multiple evidence in the bayesian belief network can be carried at in of several ways, being that each form corresponds to a similarity function in the vector model. The analysis of this correspondence is made through the functional framework. We show that the framework allows us to design new models and helps designers to modify these models to extend them with new evidence sources.