A Layered Bayesian Network Model for Document Retrieval

Abstract

Abstract. We propose a probabilistic document retrieval model based on Bayesian networks. The network is used to compute the posterior probabilities of relevance of the documents in the collection given a query. These computations can be carried out efficiently, because of the specific network topology and conditional probability tables being considered, which allow the use of a fast and exact probabilities propagation algorithm. In the initial model, only direct relationships between the terms in the glossary and the documents that contain them are considered, giving rise to a Bayesian network with two layers. Next, we consider an extended model that also includes direct relationships between documents, using a network topology with three layers. We also report the results of a set of experiments with the two models, using several standard document collections.

A New Subject-based Document Retrieval from Digital Libraries Using Vector Space Model

“…The model was represented by a network of association among concepts defining key domain entities and is extracted from a corpus of documents or from a domain knowledge base. Campos et al proposed a probabilistic model based on Bayesian network for document retrieval [11] and used the network to compute posterior probabilities for the relevance of the documents. Mohebi et al proposed a new subject-based retrieval method to retrieve all documents from a scientific digital library related to that subject.…”

A Multi-layered Bayesian Network Model for Structured Document Retrieval

“…Bayesian networks have also been successfully applied in a variety of ways within the IR environment, as an extension/modification of probabilistic IR models [6,13,16]. We shall focus on a specific BN-based retrieval model, the Bayesian Network Retrieval Model with two layers (BNR-2) [1,6], because it will be the starting point of our proposal to deal with structured documents. The set of variables V in the BNR-2 model is composed of two different sets 1 , V = T ∪ D: The set T = {T 1 , .…”

“…Structural units U ij in level j, j = l: to estimate p(u ij |pa(U ij )) we intend to use a similarity measure between two sets of terms, one associated to the whole unit U ij and the other associated to the units contained in U ij that are instantiated as relevant in the configuration pa(U ij ). More precisely, let R(pa(U ij )) = {U kj+1 ∈ P a(U ij ) | u kj+1 ∈ pa(U ij )}, and let A(U ij ) and A(R(pa(U ij ))) be the sets of terms that have been used to index U ij and the units in R(pa(U ij )), respectively 6 . In graphical terms, A(U ij ) = {T k ∈ T | T k is an ancestor of U ij } and A(R(pa(U ij ))) = {T k ∈ T | T k is an ancestor of some node in R(pa(U ij ))}.…”

“…In this paper we propose a BN model for structured document retrieval, which can be considered as an extension of a previously developed model to manage standard (non-structured) documents [1,6]. The rest of the paper is organized as follows: we begin in Section 2 with the preliminaries.…”

Implementing relevance feedback in the Bayesian Network Retrieval model

“…Nowadays, the dominant approach to managing probability within the field of Artificial Intelligence is based on the use of Bayesian networks (Pearl, 1988). They have also been successfully applied in the IR environment as an extension/modification of probabilistic IR models, giving rise to the Inference Network (Turtle, 1990; Turtle & Croft, 1990, 1991), Belief Network (Silva, 2000; Ribeiro‐Neto & Muntz, 1996), and Bayesian Network Retrieval (de Campos, Fernández‐Luna, & Huete, 2000, 2002a, 2002b; Fernández‐Luna, 2001) models, as well as other approaches (Fung & Favero, 1995; Ghazfan, Indrawan, & Srinivasan, 1996).…”