Improving Entity Linking with Graph Networks.

Entity linking aims to assign a unique identity to entities mentioned in text given a predefined Knowledge Base. Previous works address this task based on the local or global features or the combination of them. However, they are faced with the following problems: 1) For the local features based models, their decisions tend to choose entities with high external knowledge support due to the unbalanced training data and supporting score combination strategy. 2) For the global features based methods, the collective entity linking methods suffer from high computational complexity while the sequential decision model may ignore the correlation between mentions. To tackle the problem of local models, this paper proposes to leverage graph convolutional network for entity embeddings, which could integrate global semantic information and latent relation between entities. We also utilize multi-hop attention mechanism to strengthen the expression of mention context and balance the contributions of mention context and external knowledge. To tackle the problem of global methods, we put forward a global sequential inference model with graph-based search algorithm to model the coherence between mentions with low computation cost. Extensive experiments show that our model could achieve competitive results on multiple standard datasets.