Research On Relation Extraction Method Based On Recurrent Neural Network

With the rapid development of the Internet,a large amount of unstructured text data is increasing.How to structure these non-text data has become an urgent problem,and the research object of natural language processing tasks is precisely these unstructured data.In the field of natural language processing,entity relationship extraction is one of the important tasks.Entity relationship extraction tasks are of great significance for artificial intelligence and related research fields such as question answering systems,knowledge maps,and information retrieval.The relation extraction task can automatically construct the knowledge existing in the sentence by acquiring the relationship between the entity words in the sentence,and it is convenient for the researcher to construct the knowledge graph.Traditional relation extraction methods mainly use feature-based and matching-based methods,but these methods rely heavily on artificial experience and require skilled and relevant researchers to manually select features,which leads to inefficient and error-prone feature selection.With the development of deep learning in recent years,more and more deep learning methods have been applied to natural language processing tasks,and have achieved great success.The research work of this thesis mainly includes the following three aspects:(1)This paper proposes a relationship extraction model based on dual-channel self-attention.This model uses convolutional neural network and recurrent neural network to extract data features respectively,then fuse these features,and finally extract entity relationships through the fused features.This model makes full use of the respective advantages of convolutional neural networks and recurrent neural networks to make up for each other's shortcomings.Finally,the experimental results show that this model is trained and tested on the Sem Eval-2010 Task8 data set.The test results are compared with other data sets using this.The six models have all improved,and the F1 value of the comprehensive evaluation index is 1.4% to 6.3% higher than other methods,reaching 85.1%.(2)Considering the influence of entity words in the sentence on the entity relationship extraction task,by increasing the influence of named entity recognition tasks on the model parameters,thereby strengthening the influence of entity words on the relationship extraction task,this paper proposes a relationship extraction model based on the auxiliary model.This model adds the entity recognition task as an auxiliary task to the entity relationship extraction task to the model training,thereby further improving the role and influence of entity words in sentences on the entity relationship extraction task.During the test process,only the main model is used for entity relationship extraction task testing,and entity recognition tasks are no longer used.The experimental results show that under the same data set as the previous model,the F1 value of the comprehensive evaluation index obtained by the model is 1.0% higher than the previous model,reaching 86.1%.(3)To strengthen the connection between entity words and sentences,this paper proposes a relation extraction model based on selection gate network.The selection gate structure can effectively obtain the important relationship between the entity word and the context,and at the same time filter the invalid information data in the sentence,and retain the data related to the sentence semantics and the entity word and context.The experimental results show that this model is trained and tested on the Sem Eval-2010 Task8 dataset.The test results are improved compared to other 6 models using this dataset.Its comprehensive evaluation index F1 is higher than other methods 2.8% to 8.1%,reaching 86.9%.