Resorting Relevance Evidences to Cumulative Citation Recommendation for Knowledge Base Acceleration

Most knowledge bases (KBs) can hardly be kept up-to-date due to time-consuming manual maintenance. Cumulative Citation Recommendation (CCR) is a task to address this problem, whose objective is to filter relevant documents from a chronological stream corpus and then recommend them as candidate citations with certain relevance estimation to target entities in KBs. The challenge of CCR is how to accurately category the candidate documents into different relevance levels, since the boundaries between them are vague under the current definitions. To figure out the boundaries more precisely, we explore three types of relevance evidences including entities’ profiles, existing citations in KBs, and temporal signals, to supplement the definitions of relevance levels. Under the guidance of the refined definitions, we incorporate these evidences into classification and learning to rank approaches and evaluate their performance on TREC-KBA-2013 dataset. The experimental results show that all these approaches outperform the corresponding baselines. Our analysis also reveals various significances of these evidences in estimating relevance levels.