Named Entity Based Document Similarity with SVM-Based Re-ranking for Entity Linking

In this paper we present a novel approach to search a knowledge base for an entry that contains information about a named entity (NE) mention as specified within a given context. A document similarity function (NEBSim) based on NE co-occurrence has been developed to calculate the similarity between two documents given a specific NE mention in one of them. NEBsim is also used in conjunction with the traditional cosine similarity measure to learn a model for ranking. Naive Bayes and SVM classifiers are used to re-rank the retrieved documents. Our experiments, carried out on TAC-KBP 2011 data, show NEBsim achieves significant improvement in accuracy as compared with a cosine similarity approach. They also show that re-ranking using learn to rank techniques can significantly improve the accuracy at high ranks.