Research On Distant Supervision Relationship Extraction Based On Deep Reinforcement Learning

In this era of data explosion,how to quickly and efficiently from the open area or specific areas of information and technology through the analysis of the text to get effective information,has become the current is an important problem in front of people,entity relationship is automatically extract text data mining,and a core task in the process of information resource extraction.The relational extraction model based on deep learning can automatically learn semantic and relational features from massive texts.However,the model needs to rely on large-scale annotated text data for training.In order to quickly obtain large-scale training data with annotations,distant supervision automatically generates large-scale annotated data for relational extraction by aligning the knowledge base with free text.However,due to the imperfection and bias of the knowledge base,the distant supervision method will lead to noise labeling problem.In order to reduce the mislabeled noise data extracted by distant supervision relation in the training data set and improve the performance of the relational extraction model,the following two aspects are explored in this paper:1.Relation Extraction Model Based on Word ClassificationThe strategy network in the distant supervision relation extraction denoising model based on deep reinforcement learning needs to be parameterized by supervised relation extraction model.Therefore,on the basis of the relationship extraction model based on the attention mechanism,a model to enhance relationship extraction through word classification task is proposed to further strengthen the role of key words in relationship extraction and enhance the ability of relation extraction.The model is abstracted into a policy network,which is used to identify the noise data from the training data set by distant supervision relation extraction.The seed bootstraps are used to construct a glossary of relation keywords to learn relation awareness.This paper proposes a relational keyword classification task,which classifies words in sentences according to the representation of hidden layers in the neural network.The word classification task was trained with the relation extraction model,and the model parameters were shared to help the neural network identify and highlight the relation keywords with relation attributes,and to strengthen the role of the relation keywords in the relation extraction,so as to learn a new feature representation form for the relation extraction task,to improve the performance of the relation extraction model.2.Distant Supervision Relation Extraction Denoising Model Based on Deep Reinforcement LearningDistant supervision relation extraction in order to avoid the time-consuming and labored problem of annotation of training data set,a large number of training data with annotation are quickly obtained by matching free text to knowledge base,and then the model is trained with these data.However,due to the strong assumptions in the process of generating training data,there are a large number of mislabeled noise data in the training data.Using data sets with noise data to train the model will inevitably affect its performance.In this paper,the deep reinforcement learning method is used to remove the noise data mislabeled by distant supervision.The model consists of two modules: agent and relation extraction network.An agent with autonomous learning ability based on strategy network is designed to recognize noise data and remove noise data.The agent is initialized through pre-training,so that it has a certain ability to identify noise data,and its performance is continuously improved through subsequent interaction with the relationship extraction network.Finally,a data set with relatively less noise data is obtained.Experimental results show that this method is effective for noise data removal.