System for collective entity disambiguation

Abstract

We present an approach and a system for collective disambiguation of entity mentions occurring in natural language text. Given an input text, the system spots mentions and their candidate entities. Candidate entities across all mentions are jointly modeled as binary nodes in a Markov Random Field. Their edges correspond to the joint signal between pairs of entities. This facilitates collective disambiguation of the mentions achieved by performing MAP inference on the MRF in a binary label space. Our model also allows for a natural treatment of mentions that either have no entity attached or have more than one attachments. By restricting cliques to nodes and edges and with a submodularity assumption on their potentials, we get an inference problem that is efficiently solved using graph min cut.