An Approach to Collective Entity Linking

Entity linking is the task of disambiguating entities in unstructured text by linking them to an entity in a catalog. Several collective entity linking approaches exist that attempt to collectively disambiguate all mentions in the text by leveraging both local mention-entity context and global entity-entity relatedness. However, the complexity of these models makes it unfeasible to employ exact inference techniques and jointly train the local and global feature weights. In this work we present a collective disambiguation model, that, under suitable assumptions makes efficient implementation of exact MAP inference possible. We also present an efficient approach to train the local and global features of this model and implement it in an interactive entity linking system. The system receives human feedback on a document collection and progressively trains the underlying disambiguation model.