Multi Label Entity Relation Extraction Research Based On Deep Learning

Entity relation extraction is one of the most critical basic tasks in the field of Natural Language Processing(NLP).It is necessary to automatically extract potentially complex semantic relationships from a large amount of unstructured network text data.Multi-label entity relation extraction is a more complex task in entity relation extraction,which refers to extracting the relation between multiple entities in a sentence containing multiple entities.Extracting the relation between multiple entities from a single sentence and multiple entities is more complicated than extracting the relation between two entities from a single sentence and double entities.At present,scholars apply a large number of deep learning network models to entity relation extraction tasks in the field of natural language processing,such as some improved models based on Convolutional Neural Network(CNN)and Recurrent Neural Network(Recurrent Neural Network,The improved model of RNN has achieved good results in entity relation extraction.But these models have only achieved good results in the extraction of single-sentence and dual-entity relation.Due to the local calculation of the convolutional neural network model and the characteristics of multi-label data set relation overlap and feature discreteness,the experimental results of applying this type of model to multi-label entity relation extraction are not ideal.In response to the above problems,the following research is carried out in the research of multi-label entity relation extraction:(1)The experimental training data set is the crawled batch of text data in the field of humanistic information.The capsule network model is used for feature aggregation,and attention is added to the dynamic routing algorithm of the traditional capsule network to transfer information from low-level capsules to high-level capsules.Mechanism,improved the original capsule network model.In the process of information transmission,the improved model can more accurately enlarge the transmission of effective features,reduce or even ignore the transmission of useless feature information.By comparing the experimental results of the model,the effectiveness of the improved model is verified.(2)In order to reduce the impact of the original noise data in the dataset on the entire model,a sentence filter model is added,and a joint model that shares some parameters in sentence filtering and relation extraction is used for processing.First,perform one-step filtering and noise reduction processing on the noisy data in the data set,and then complete entity relationship extraction from the noise-reduced data.The capsule network model with improved dynamic routing algorithm is used in the task of relation extraction.The two models share model parameters.Joint training updates the shared parameters in each iteration of the entity relation extraction model,and then uses them as the input parameters of the sentence filter to select better sentences,so that the two models can influence and optimize each other during the training process.By comparing the experimental results with other models,the superiority of the combined model is verified.