Composite Documents and Uncertain Inference

The composite nature of multimedia documents requires a more powerful knowledge representation for indexing than the pure set of terms. Object-oriented data modeling is a widely known and accepted approach to represent knowledge. [Meghini et al. 93] introduces the usage of object-oriented principles in combination with logic for improving information retrieval. This workshop contribution presents ideas to combine uncertain inference with object-oriented modeling in order to achieve a suitable model for multimedia information retrieval. This work introduces the probabilistic extension of the model presented in [Rölleke & Fuhr 96]. Figure 1 depicts the major issues of the model. The documents d1 andd2 consists of words and sections. The square brackets indicate the composite (aggregated) structure of documents. This concept of aggregation allows for reflecting the composite nature of multimedia documents appropriately. In addition, it is suitable for modeling retrieval among distributed environments as indicated by the two databases db1 and db2. We consider databases, documents, and sections as contexts which define a local frame for a logical program. A logical program is a set of facts and rules for defining a set of propositions. Following the object-oriented principles, we consider three types of propositions, namely classification, generalization, and roles. Classification serves to group objects within certain classes (concepts). For example, document d1 states that object peter is an instance of class sailor. Generalization serves to define class hierarchies. For example, every picture is a document. Roles represent the relationships between objects. For example, mary is the author of d1. In addition, our model allows to use a predicate with arity zero (e. g. sailing). This corresponds to the classical set of terms approach and provides the familiar way for describing the content of documents.