Approximate answers in intelligent systems
Berthier A. N. Ribeiro · 1995
This dissertation investigates the problem of approximate query answering in the context of bibliographic and relational database systems. A bibliographic database is a collection of documents (i.e., books, proceedings, reports) indexed by keywords. The user searches for the documents of his interest by specifying a set of keywords. Thus, a user's query request is inherently vague. Bibliographic databases are commonly referred to in the literature as information retrieval (IR) systems. A relational database system is a collection of relations. The user searches for the tuples of his interest by specifying a query in SQL--a high level query language which is precise. However, many modern applications (e.g., geographical information systems and medical applications) require the system to implement some notion of approximation to facilitate the user's querying process. A relational database extended to allow approximate queries is commonly referred to in the literature as a cooperative database system. Our work proposes the application of Bayesian belief networks for generating approximate answers in the context of IR and cooperative database systems. Further, we also propose a fuzzy set model for cooperative databases. The network model we propose for IR is an alternative to the Turtle & Croft Inference Network Model and overcomes important drawbacks of that model. For instance, our model subsumes all classic models in IR while theirs does not. Further, our model can be extended with information from responses to previous query requests to yield improved retrieval performance. We also propose a second belief network model which can include information about relevant documents provided by the user--a process usually referred to as user relevance feedback. We show that this network model can be extended with information from previous feedback cycles to yield improved retrieval performance when compared to strategies based on the Rochio formulation. In the context of cooperative databases, we propose a belief network model which can be successfully used to merge semantic metric distances, complex queries, and a ranking strategy in consistent fashion. Through examples, we illustrate the main advantages of this model. The fuzzy set model we propose for cooperative databases is based on two basic principles that can be easily grasped by the user. As a result, the user is able to mentally mimic the system behavior to gain greater understanding of the ranking.