| Since 2012,Google Knowledge Graph technology has attracted extensive attention from researchers in various fields.In essence,Knowledge Graph is a network semantic knowledge base formed by the connection of entity nodes.Its original intention is to enhance users' search experience and enhance the ability of search engines.However,with the in-depth research of researchers,Knowledge Graph technology has penetrated into various research fields.With its powerful open organizing and semantic processing ability,it has laid a foundation for the intelligent information organization and application in the Internet era.Knowledge Graph aims to clearly describe the relationship between entities,that is,entity relationship pair is the basic unit of knowledge Graph.Therefore,knowledge extraction,as the foundation of knowledge Graph,has become an important research content in the field of natural language processing in recent years.Entity relationship extraction has two work modes;the traditional way of working is similar to Pipelined: For the input sentence sequence,firstly,the entity is recognized,and then the recognized entities are combined by using a certain rule,and then the relationship is extracted(classified),thereby obtaining an entity relationship pair composed of the entities and the corresponding relationship type.The second is the joint extraction work mode,which completes the named entity recognition and relationship extraction tasks simultaneously through a single model.This paper takes ACE05 data and NYT open source dataset as the research object,combines Deeply Neural Network and Attention Mechanism and Distant Supervision,proposes two new joint entity relationship extraction models,to complete the named entity recognition and relationship extraction tasks simultaneously.The main work of this paper includes the following two parts:(1)the traditional entity relationship extraction model leads to error accumulation and information redundancy,the relationship between the two subtasks is not considered in the actual model training.To solve these problems,this paper proposes a joint entity relationship extraction model based on Shared underlying coding and single-double Attention layer mechanism.This scheme preserves the Shared underlying neural network coding approach to learning,and adds an E2E-Attention layer to calculate the interaction among the entity words,and an E2R-Attention layer to calculate the interaction between the entity word and the relationship word.In relationship extraction module,we adopted the Bi-LSTM network to consider the influence which location of the entities words to relationship classification.By sharing the underlying coding layer,and using the single-double Attention layer which nonsimultaneous work to enhance the dependence of entity words and relationship words in the knowledge extraction task.Then the back-propagation algorithm is used to update the shared underlying parameters,thereby realizing two sub-tasks' Joint learning.This method enhances the adaptability of the model to the actual experimental data,improves the accuracy and comprehensive performance of the entity relationship extraction task.(2)A semi-supervised joint entity relationship extraction model based on Sequence labeling Strategy and Distant Supervision is proposed.This scheme is based on the sequence labeling strategy,which converts the input sentence into a 1-hot word embedding through word embedding learning,and then passes it to the LSTM encoding layer and the information enhance layer.Among them,the information enhance layer mainly includes an Attention layer for calculating the dependency relationship between the entity relationship pairs,and an enhance weight layer for the entity relationship pair that obtained by using the distant supervision.We use these two sub-weighting modules to reinforce the model's attention on the entity words and potential relationship types,thereby transforming the entity relationship extraction tasks into end-to-end sequence labeling tasks,realize the joint entity relationship extraction task in time and space.the model's dependence on manual labeling data is greatly reduced to a certain extent,and it plays a very important role in the construction of vertical domain knowledge graph. |
