Neural Collective Entity Linking Based on Recurrent Random Walk Network\n Learning

Abstract

Benefiting from the excellent ability of neural networks on learning semantic\nrepresentations, existing studies for entity linking (EL) have resorted to\nneural networks to exploit both the local mention-to-entity compatibility and\nthe global interdependence between different EL decisions for target entity\ndisambiguation. However, most neural collective EL methods depend entirely upon\nneural networks to automatically model the semantic dependencies between\ndifferent EL decisions, which lack of the guidance from external knowledge. In\nthis paper, we propose a novel end-to-end neural network with recurrent\nrandom-walk layers for collective EL, which introduces external knowledge to\nmodel the semantic interdependence between different EL decisions.\nSpecifically, we first establish a model based on local context features, and\nthen stack random-walk layers to reinforce the evidence for related EL\ndecisions into high-probability decisions, where the semantic interdependence\nbetween candidate entities is mainly induced from an external knowledge base.\nFinally, a semantic regularizer that preserves the collective EL decisions\nconsistency is incorporated into the conventional objective function, so that\nthe external knowledge base can be fully exploited in collective EL decisions.\nExperimental results and in-depth analysis on various datasets show that our\nmodel achieves better performance than other state-of-the-art models. Our code\nand data are released at \\url{https://github.com/DeepLearnXMU/RRWEL}.\n