Research And Implementation Of Relation Extraction Based On PLSTM-CNN And Shared Representation Generator

Entity relation extraction is an important task in natural language processing.Because the traditional supervised classification method often needs lots of inefficient manual labeling work,relation extraction based on distant supervision has become a new research hotspot.There is a strong hypothesis in distant supervision that if two entities have some kinds of relationship,this relationship exists in any sentences containing both two entities at the same time.However~,this hypothesis is not always true in practice,which leads to more noises in the data set of automatic annotation.How to alleviate the noises is a difficult problem to be solved urgently at present.This paper improves the performance of relation extraction based on multi-instances learning from two aspects:sentence vector representation and bag vector representation.In sentence vector representation,BiLSTM is introduced to learn more association features to improve the piecewise convolution neural network which ignores the problem of semantic relevance.In bag vector representation,a new shared representation generator is proposed to transform the feature space,which maps sentences from the original semantic space to the feature space related to the semantics of the target relationship,so as to filter out the expression of irrelevant noises.Different from the mainstream attention mechanism,the proposed method has stronger non-linear fitting ability and can extract more common features of instances.In addition,this paper introduces the extra generator loss to enhance the performance of generator.Our method has good scalability and the implementation of shared representation generator is customizable and extensible.Finally,five groups of experiments on two different data sets are carried out.The results show that the performance of the method proposed in this paper is significantly improved compared with that of baseline,and it can effectively alleviate the problem of noise interference in distant supervision.