A Collective Entity Linking Method Based on Graph Embedding Algorithm

Aiming at the poor effect of traditional entity linking method combining context co-occurrence and objective knowledge, this paper proposes a collective entity linking framework based on graph embedding algorithm Node2vec. The framework first uses BiLSTM and CRF to do named entity recognition over a text, aiming to obtain entities related to mention in the text. Then, the Node2vec is used on the sub-graph of the knowledge graph for all entities corresponding to the candidate entity sets and the recognized entities to obtain the vector representation between different entities. The similarity between each entity and the text within a certain range is calculated, and the degree between each pair of entity elements is obtained. The most similar pair is selected as the target entity of the link. Experimental verification shows that F1 in this framework reaches 85.96% on NLPCC2014 data set, and the improved Node2vec algorithm performs better than some other graph embedding algorithms in this framework.