Research On Joint Extraction Of Entity Relationship Based On Graph Neural Network

Entity relationship extraction is an effective technology for mining network value information.By extracting key data from unstructured natural language texts for structured representation,it provides a data basis for downstream tasks such as search question and answer and knowledge graph.At this stage,entity relationship extraction is mostly the traditional pipeline method and joint extraction method.The traditional pipeline method cascades two subtasks of named entity recognition and relationship extraction.The model is simple but the extraction ability is lacking,and the cascade method has problems such as error accumulation and lack of subtask relevance.The joint extraction method is the unified modeling of two subtasks,and the end-to-end model training is completed through one model,which solves the model structure defects of the Pipeline method,but the model extraction ability is not mature.In the entity relationship extraction task,there are multiple overlapping problems between entities and entities corresponding to multiple relationship categories,and optimizing the overlap problem has become the key to the modeling of joint extraction methods.Under the above background,this subject proposes a joint extraction model based on graph neural network based on in-depth study of entity relationship extraction and deep learning models.Firstly,on the basis of the data processing method of the pointer labeling structure,the dependency syntax analysis technology is introduced to provide the global features of the text,and the rules for transforming the dependency syntax analysis graph structure into the dependency adjacency matrix are designed,and the original dependency syntax analysis feature is converted into the graph matrix feature.Then introduce graph attention and other network mining graph matrix features and text features.In order to make effective use of the dependency syntactic structure information,this topic introduces the self-attention mechanism into the graph attention network and proposes the SLGAT(self-learning graph attention network)model,which gives the graph attention network self-learning ability and autonomously adjusts the adjacency matrix that is beneficial to the model.Weight,thereby improving the effect of the model and optimizing the overlap relationship problem.Finally,considering the applicability of the feature extractor and the influence of the text structure and other factors,this topic combines the serialization model bidirectional long-term short-term memory network and SLGAT to propose a Bi LSTM-SLGAT entity relationship joint extraction model.Related experiments show that the Bi LSTM-SLGAT model is optimally improved by about 6%.In addition,in order to enhance the model’s ability to express text features and further enhance the effect of model joint extraction,this topic introduces word vector technology and model fusion.Among them,the Fine-tuning of the BERT word vector has been improved accordingly,and the multi-dimensional output results of the BERT original word vector are stacked to further improve the text feature representation.The experimental results show that the model of fusion word vector is optimized by about 3%.To sum up,this subject has studied the joint extraction model based on graph attention network and its fusion word vector for the modeling method,model effect and entity relationship overlap in entity relationship extraction technology.The experimental analysis and evaluation at the end of the subject show that This method can effectively improve the effect of entity relationship joint extraction.Finally,I hope that this topic can provide some reference and help for the joint extraction of entity relationships.