Research On Entity Relation Extraction Technology Based On Deep Learning

As we can see,a great deal of information is displayed in unstructured electronic documents with the coming of the Internet and Big Data era.It has became a burning issue to structure these text data with high quality and speed,and Information Extraction(IE)has emerged under this background.The research of IE is designed to extract high-quality and structured available data from complex and redundant texts at low cost,which greatly promotes the development of natural language processing.Relation extraction(RE)forms the regular triple of entity relations through finding the relation type information among entities,so as to construct the unstructured text into structured text.As the core tasks of IE,it has great impact on artificial intelligence such as knowledge graph,recommendation retrieval and intelligent question and answer.RE tasks have adapted more and more deep learning methods,and have achieved relatively fruitful results over the years.However,these methods based on deep neural networks still have many shortcomings.In order to obtain the global information of the text,most models based on convolutional neural networks can only stack multiple layers and complicate the model to increase the receptive field.While models based on recurrent neural networks can learn long-distance dependencies,but they are difficult to achieve parallelization.Regarding the above points,two improved RE models are introduced in the paper,one is the RE model based on graph convolutional network(GCN),and the other is the RE model based on BERT representation.The material work of the paper sets out below:(1)This thesis Designs and implements a weighted convolutional neural network(WGCN)model for RE.For dependency relation extraction task,the GCN is used to ob-tain dependency structure information of dependency syntax tree,and the idea of logical adjacency matrix replacing ordinary adjacency matrix is used to improve the graph con-volution network.Combining with the attention mechanism of entities,it focuses on the relationship prediction of global information highly associated with entities.Compared with the ordinary GCN model,it is proved that the WGCN model is effective in the task of RE.(2)This thesis Designs and implements a method based on BERT representation and fusion of latent entity types for RE tasks.For the end-to-end RE model,based on the“pre-training+fine-tuning”training mode,the BERT is adopted to obtain the dynamic word embedding of the text.Firstly,from the perspective of the entity itself,we learn the potential type representation of the entity and supplement the features.Then,the key in-formation of the text is captured by the entity text attention mechanism for RE.The method can enhance the robustness of RE model subject to the experimental results.