Deep Learning Based Entity Linking:Method And System

Entity linking aims to map entity mentions in text to corresponding correct entities in an existing knowledge base.As a subtask of natural language processing,it is one of the key techniques for eliminating the ambiguity in natural language and help computers understand natural language.The current mainstream entity linking methods construct a subgraph for the candidate entities of the entity mentions in the document,and then combine the global features extracted from the word graph with the local features to sort the candidate entities.This kind of methods make use of the relationship information between candidate entities,but they are computationally intensive and do not make use of the semantic information of the text between two entity mentions in the document.For this reason,this thesis adopts deep learning to study the impact of the semantic information of the text between two entity mentions on the correct candidate entity ranking in the model.Based on this,this thesis proposes a learning method for joint training of word vectors and entity vectors and a disambiguation model enhanced by the semantic information of the text between two entity mentions.It finally develops an entity linking system,which can display the linking results in a visual way.The specific contributions of this thesis are as follows:1.It proposes a new method for jointly training word vectors and entity vectors.In deep learning-based entity linking methods,both words and entities need to be represented using vectors,Moreover,the context of the mention and the candidate entity should interact to obtain the similarity between the mention and the candidate entity,this method takes into account the features of the previous joint learning method with complex parameters and difficult to adjust parameters,which improves the efficiency of the joint training method.It embeds words and entities naturally in the same vector space.The experimental results show that the proposed method can well express the structural information and semantic information between entities in the knowledge base.The generated entity vector achieved the best level on the NDCG@1 indicator of entity correlation and entity analogy reasoning experiment.2.It proposes a joint disambiguation model enhanced with semantic information of the text between entity mentions.Upon observation,it is found that the context between entity mentions in the document has a certain semantic similarity to the entities corresponding to the mentions.Based on the entity vectors generated previously,a joint disambiguation model with enhanced semantic relations between entity mentions is proposed in this thesis.This model uses deep learning technology to model the text between entity mentions and then uses contextual semantic information to help the model infer global inference.By combining with the local model proposed in this thesis for training,this thesis obtains results comparable to the current best results in the entity linking dataset.3.It develops an entity linking system.The front-end interface receives natural language text submitted by the user.After sorting the candidate entities through the entity linking model,the system marks the candidate entity with the highest confidence as the correct entity.The mentions in the text are then converted into hyperlinks to the pages corresponding to the correct entities.The background also includes a knowledge base information update mechanism to ensure the accuracy and real-time of the linking results.The system verifies the effectiveness of the deep learning-based entity linking model.At the same time,the system is also applied to scientific research topics such as "Knowledge base construction based on open network data" and other downstream tasks.Through the above work,the entity link model we have obtained has achieved a similar level to the current best results on related datasets,which proves the effectiveness of our work.This work provides new ideas for collaborative entity linking methods and further promotes the development of entity linking research methods.