Novelty Detection in Knowledge Base Acceleration
Master thesis
Permanent lenke
http://hdl.handle.net/11250/253279Utgivelsesdato
2013Metadata
Vis full innførselSamlinger
Sammendrag
Knowledge bases provide the users of the World Wide Web with a vast amount of structured information. They are meant to represent what we know about the world the way it is today. Therefore, every time something happens, knowledge bases need to be updated according to the new happening. A knowledge base is most often organized around entities and their relations. Entities represent an object in the real world, such as religions, persons or places, and a relation is a connection between two entities. Today, the process of updating knowledge bases is purely done by humans, who unfortunately are not able to keep up with everything that happen in the world. In order to make this job easier, systems for doing Knowledge base acceleration, KBA, are proposed. They are meant to, given a stream of news, pick out what is relevant updates for the different entities in a knowledge base. To make the most of such a system, and to make sure that it only return news that provide useful information to the content managers, it should only return news that contain \textit{new} information, that is, it should perform novelty detection. This thesis explore the properties a KBA system need to fulfil in order to solve the task it is supposed to as good as possible. It argues that a KBA system need to include novelty detection to be useful, and present a prototype for novelty detection in a KBA system. The prototype is implemented using different approaches to novelty detection, and compare these.