Research Of Collective Entity Linking Based On Joint Embedding Of Word And Entity

Unstructured natural language text is often ambiguous,especially named entities.A named entity can have multiple mentions,and a mention can also represent several different named entities.The entity linking task is to link the corresponding entities in the unstructured text to the structured knowledge base,which helps to mine the information of the original noisy text and realize the vision of the Semantic Web at an early date.The previous researches on entity linking task are mainly based on statistical models,which rely on artificially defined features about text and entities.These features are usually defined by experts in the field,and may not capture all relevant statistical dependencies and interactions.Neural network models can automatically learn features,and have outstanding advantages in semantics coding and data sparsity processing.Therefore,this paper designs an entity link model based on neural network to automatically learn features and their combinations.In this paper,we first construct an extended Skip-gram model to embed words and entities jointly in a low-dimensional vector space and use it as input to entity link model.The embeddings of words and entities compress the semantic information and avoid the manual extracting features.Then,we build a neural network model to automatically learn the two features about mention-entity semantic matching and entity pair consistency.We model the collective entity linking task as a conditional random field to combine these two features.Finally,we conducted experiments to test the models.The experimental results show that the text-entity joint embedding model in our paper can learn the semantic information of words and entities well,and the entity link model achieves a microscopic F1 score of 91.87% on the AIDA data set.