OntoHop: An information filtering agent using hopfield nets and ontologies

The size of the Web and its dynamic nature in addition to the fact that stored documents are written in natural language, and therefore intended to be read by people and not to be processed by computers, present major challenges to build automatic personalized information filtering systems. This article presents the architecture of an information filtering agent based on an implementation of a Hopfield neural network (HNN). Network nodes (neurons) represent relevant terms in the domain of interest and neuronal links represent asymmetric probabilities of term co-occurrences in the domain, or the relevance weight between a pair of terms. Relevant terms are automatically derived from a corpus related to the domain of interest using automatic indexing and an ontology. Co-occurrence probabilities are computed by a cluster function that produces asymmetric links between terms. At the moment of document filtering, input neurons are activated on the basis of the presence of terms in the document that are identical or semantically similar to the terms stored in the net. The semantic similarity between terms is calculated using a hierarchical ontology that describes concepts that exist in the domain of interest. Experiments conducted to evaluate the precision and recall of the agent with and without the use of ontologies show that ontology use tends to favor recall over precision. The degree to which this bias occurs can be adjusted by setting the minimum level of similarity required to consider a document and a network term similar.