Key Research Of Entity Linking Based On Incremental Embedding

Entity linking is the process that pairs a string representing an entity in a natural language with its corresponding entity in a knowledge base.Nowadays,richer network data and closer connection of daily life and the Internet makes the entity linking an important task in understanding and processing the Internet information.The difficulty of the entity linking is that the same entity may have multiple forms of appearance,the same string may also represent multiple entities.In the existing research,entity linking systems use morphological features such as word form and noun co-occurrence,and the accuracy rate has reached the bottleneck.Using deep learning to understand the deep semantic features of the text is expected to break through this bottleneck.Existing entity-word co-embedding method requires large training data,and can not flexibly expand the collection of entities.The resulting entity embedding representation and word embedding are tightly coupled and are disjoint from other embedded research methods.In this paper,we propose a method of entity incremental embedding,which extends the model based on existing word vectors and language models,and obtains a embedding representation of entities.In particular,the initial values setting algorithm and training method of the extended model are designed to make the model convergent.Candidate ranking method based on entity embedding and candidate ranking method of combinatorial model are designed for the entity linking task.Experimental results on the TAC2010 test set show that the method of incremental embedding is feasible and effective,and a high-quality embedding representation is achieved.It can meet the requirements of small batch embedding and multiple incremental embedding turns.Candidate ranking method based on entity embedding and candidate ranking method of combinatorial model are 80% and 82.8%,respectively,which are competitive against benchmarks of similar methods.