Research On The Deep Learning Model Of Relation Extraction With Syntactic Information

With the development of human society becoming more intelligent,data such as video,audio,and text on the Internet is increasing rapidly,and information extraction technology is emerging.Entity relationship extraction,as one of the core technologies of information extraction,aims to predict the semantic relationship between two or more target entities in a text,which is of great significance in the fields of knowledge base construction and question answering system.Compared with the traditional method based on feature and kernel function,the relation extraction model based on deep learning has obvious advantages in both classification accuracy and feature extraction depth,and the relation extraction model based on dependency analysis tree and attention mechanism has been gradually proposed.In this thesis,syntactic information is introduced to obtain the local semantic features of the entity context,and on the basis of the pre-trained language model,the features of each granularity are fully utilized to obtain more abundant semantic information for relation classification.The main research work of this thesis is as follows:Different from the conventional text classification task,the relation extraction task needs to emphasize the semantic information of entity pairs and entity context.Therefore,this thesis proposes a relation extraction model based on the attention mechanism of dependency analysis to obtain the local semantic features of the entity context.The model is able to capture the most significant unary grammar,binary grammar fields and local features of relation classification,take into account the interdependence between each word and the target entity to highlight the target entity,and eliminate the influence of noisy words when the target entity is far away.Previous fine-tuning methods for single sentence classification tasks ignore the rich semantic knowledge contained in the intermediate layer of the pre-trained language model,this thesis proposes a relation extraction model based on the semantic information of the intermediate layer of the pre-trained language model,aims to explore how to improve the fine-tuning of the pre-trained language model by utilizing the intermediate layer of the pre-trained language model.In a specific natural language processing task,different layers of the pre-trained language model make different contributions to the semantic information obtained,so how to comprehensively utilize the intermediate layer to obtain more abundant semantic information is also a crucial problem.In view of the limitation of relation representation caused by the use of single granularity feature of sentence in previous models,a relation extraction model based on multi-granularity feature is proposed in this thesis.In this model,the single sentence sequence features are replaced by multi-granularity features,that is,sentence level features,target entity features and fine-grained features based on dependency analysis information are combined features for relation classification.The features of the model design are integrated with shallow features,local fine-grained features and global coarse-grained features to enrich the final feature representation for classification.In this thesis,we conducted experiments on two datasets,Semeval-2010 task 8 and KBP37,and the experimental results show that this method can improve the accuracy of semantic feature representation and entity relationship extraction.