Research And Implementation Of Relation Extraction Based On Similarity

With the explosive growth of network resources,it is difficult for users to search for the relevant information they want quickly and effectively.Entity relation extraction enables people to extract the relationship between two entities from a large amount of text information,thus structuring the information.Relation extraction is a basic task of large NLP tasks such as knowledge graph and search engine,which is of great significance to NLP research.Compared with traditional methods such as kernel function,neural network model has obvious advantages in mining potential relationships,and has become the mainstream model to deal with relation extraction tasks.However,the neural network model also has some shortcomings in relation extraction task.For example,the point estimation parameters of the model are fixed after training,which leads to the unsatisfactory extraction ability of the model in complex sentences.Besides,the relation extraction model is easily disturbed by the similarity relation,which leads to the low classification accuracy.In view of the above problems,this thesis makes an in-depth study.1.Aiming at the problem that relation classification in relation extraction task is easily disturbed by similarity relation,this thesis proposes a loss function based on similarity relation to eliminate the influence of similarity relation in relation classification and improve the accuracy of relation classification.Firstly,the model measures the similarity of the relationships between entities,and selects the similarity relationship by using the threshold value.Then,the Logistic loss function of the similarity relationship is combined to eliminate the interference of the similarity relationship to the relationship classification task and improve the accuracy of classification.Experimental results on Wiki and NYT datasets show that compared with the original model,the model with Logistic loss function combined with similarity relationship has higher value.The effectiveness of the improved method is proved.2.In order to solve the problem that the point estimation parameter value of traditional neural network is solidified after model training,and the learning features are too rigid,which leads to the weak generalization ability of the model,this thesis introduces Bayesian long and short term memory network into the model.This improved method introduces a prior probability distribution to the parameters,and learns the posterior probability in the model training,so that the parameter values originally learned by the neural network become the parameter probability distribution.The probability distribution of parameters makes the model retain a certain confidence interval when it encounters different expressions of sentences in relation extraction task,which improves the extraction ability of the model.Experiments on the TACRED dataset demonstrate that the improved relational extraction model with Bayesian long-short memory network has higher values than other models.The effectiveness of the improved method is proved.3.This thesis designs and implements a relational question-answering prototype system based on similarity relation.The improved model is applied to question answering system to provide data support for search engine,knowledge graph and other tasks.