Design And Implementation Of Multi-Source Domain Relation Extraction Model Based On Reinforcement Learning

Relation extraction is one of the important tasks of natural language processing and understanding.It is widely used and can provide services for knowledge graph,question answering system and social network.The traditional model mainly relies on a large amount of high-quality labeled data,which consumes a lot of human and financial resources.Another promising way to expand training data is through distant supervision algorithm.The principle is that if an entity pair expresses a relation in the knowledge base,then all the data that contains the entity pair expresses the relation.In theory,this makes relation extraction a lot easier.However,the label generated by distant supervision is noisy,and the direct use of distant supervision labels will lead to poor model effect.In addition,another factor that affects the effect of relation extraction model is the quality of feature vectors.Although the feature vectors set by hand are easy to generate,they are not conducive to the classification of positive and negative samples.Therefore,it is necessary to relearn a feature vector that is conducive to improving the effect of relation extraction model.In order to solve the above problems,this paper uses the reinforcement learning method to reduce the noise caused by distant supervision,and uses the optimized label to retrain to obtain the feature vectors that are more conducive to the classification of positive and negative classes,and uses the optimized label and feature vector training to obtain a relation extraction classifier with higher classification accuracy.At first,a label noise reduction method based on reinforcement learning is proposed to solve the problem of label noise generated by distant supervision.Then,in view of the correlation between source domain and target domain data,a feature optimization method based on domain adversarial neural network is proposed,which takes into account the domain invariant features of source domain and target domain,and can learn the feature vectors more conducive to classification.Finally,we use the optimized feature vector and label training to obtain the final relational classification model.In this paper,the optimized results are compared with the results of ground truth labels and other relation extraction models.The experimental results show that the model can deal with the relational classification problem of noisy data well,and can complete the domain adaptive task well.At the end of this paper,a relation extraction model training system based on the above method is implemented.This system realizes the fully automated process of data processing,model training and result presentation,and can interact with users to adjust model parameters to adapt to different training tasks.