Multi-step classification approaches to cumulative citation recommendation

Abstract

Knowledge bases have become indispensable sources of information. It is therefore critical that they rely on the latest information available and get updated every time new facts surface. Knowledge base acceleration (KBA) systems seek to help humans expand knowledge bases like Wikipedia by automatically recommending edits based on incoming content streams. A core step in this process is that of identifying relevant content, i.e., filtering documents that would imply modifications to the attributes or relations of a given target entity. We propose two multi-step classification approaches for this task that consist of two and three binary classification steps, respectively. Both methods share the same initial component, which is concerned with the identification of entity mentions in documents, while subsequent steps involve identification of documents being relevant and/or central to a given entity. Using the evaluation platform of the TREC 2012 KBA track and a rich feature set developed for this particular task, we show that both approaches deliver state-of-the-art performance.