Filtering Documents over Time for Evolving Topics The University of Amsterdam at TREC 2013 KBA CCR

In this paper we describe the University of Amsterdam’s approach to the TREC 2013 Knowledge Base Acceleration (KBA) Cumulative Citation Recommendation (CCR) track. The task is to filter a stream of documents for documents relevant to a given set of entities. We model the task as a multi-class classification task. Entities may evolve over time and the classifier should be able to adapt to these changes at runtime. To achieve this, the clas-sifier performs online self-learning, i.e., learning only from the examples it is most confi-dent about, based on a confidence score it produces for every prediction it makes.