Research on Chinese Entity Linking Based on Multi-Granularity Text Representation and Graph Neural Network

Considering the problems of Chinese entity linking in named entity recognition, entity disambiguation, and feature dimension reduction, this paper proposes a Chinese entity linking based on multi-granularity text representation and graph neural network. First, a text representation is constructed at multiple granularities according to the BERT model to obtain character-level, word-level, sentence-level, and segment-level information about the entity candidates. Second, a feature dimension reduction strategy based on principal component analysis is proposed to eliminate the ambiguity between entity candidates. Third, based on the topological features of the graph neural network, a graph neural network model is constructed taking into account the entity context and knowledge-based influences that describe the relationships between entities in more detail. Finally, a Chinese entity linking system is built based on multi-granularity text representation and graph neural network algorithm. Simulation and experiments prove the effectiveness of the entity linking proposed in this paper.