Research On Entity Relationship Extraction Algorithm Based On Deep Learning

At present,more and more documents are presented on the Internet in the form of digital resources.The content of these documents is various and the structure is chaotic,which makes it very difficult to quickly extract effective information from these documents.Entity relation extraction,as a key technology to realize information structure,can solve this problem well,and plays a very important role in building knowledge map,intelligent question answering system and natural language generation.Relation extraction based on deep learning has a relatively good effect in relation extraction.However,data preprocessing based on deep learning algorithms adds more and more additional features obtained by natural language processing tools.The errors contained in these additional features will be accumulated in the algorithm,thereby reducing the effectiveness and effectiveness of sentence feature information.Utilization rate.Moreover,most of the algorithms use a single type of neural network for feature extraction,which leads to relatively limited features extracted by the algorithm.Aiming at these two problems,this thesis proposes new data preprocessing methods and relation extraction algorithms to improve.The specific research content is as follows:(1)In order to reduce the accumulation of a large number of errors caused by the use of too many additional features in data preprocessing,this thesis uses word vector features,location vector features,entity identification features,and sentence semantic dependency trees as data preprocessing methods.The semantic dependency tree is the only one.Additional features.(2)In order to analyze and verify the specific impact of the above data preprocessing methods on relation extraction,this thesis proposes an entity relationship extraction algorithm Re-MCNN based on multi-core convolutional neural networks and an entity relationship extraction algorithm Re-BLSTM based on bidirectional long and short-term memory networks.These two algorithms analyze the impact of the above-mentioned data preprocessing methods on the performance of relation extraction from the perspective of adjacent word features and overall sentence features.The experimental results of the algorithm show that the addition of entity identification features and semantic dependency tree to the input features improves entity relationship extraction to a certain extent,and proves that it is effective to improve the problem of a large number of errors caused by too many additional features.(3)In order to improve the problem of single neural network feature extraction,a new entity relation extraction algorithm Re-CLSTM is proposed based on Re-MCNN and Re-BLSTM.The basic idea is to use multi-core convolution neural network to extract the features of adjacent words and bidirectional long short term memory network to extract the whole sentence features.Then,the attention mechanism is used to fuse the adjacent word features with the whole sentence features.Finally,the fused features are input into the classifier to get the sentence relationship.Experimental results of Re-CLSTM algorithm show that the problem of insufficient feature extraction of single network can be improved by using two neural network fusion methods for relation extraction,and then the result of relation extraction can be improved.