A Submodular Approach for Reference Recommendation

Abstract

Choosing appropriate references for a given topic is an important, yet challenging task. The pool of potential candidates is typically very large, in the order of tens of thousands, and growing by the day. For this reason, this paper proposes an approach for automatically providing a reference list for a given abstract. The approach is based on an original submodular inference function which balances relevance, coverage and diversity in the reference list. Experiments are carried out using an ACL corpus as a source for the references and evaluated in terms of precision-recall, MAP and MRR. The results show the remarkable comparative performance of the proposed approach.