A Relation Extraction Method Based On Channelattention Mechanism And Cost-sensitive Learning

Facing the massive amount of information on the Internet,information extraction techniques have become an indispensable tool for obtaining effective information in the field of natural language processing.Among them,relation extraction,as a key sub-task of information extraction,aims to extract semantic relationships between entity pairs in sentences,which plays an important role in text semantic understanding and knowledge graph construction.Focusing on the problems of inadequate extraction and representation of semantic features of text and low performance of prediction models under the scenario of unbalanced relation categories,the following theoretical and experimental studies are conducted to improve the effect of relation extraction models in this paper.Firstly,in order to extract semantic features of text better and enhance or suppress the importance of each convolutional channel feature precisely,the convolutional feature enhancement module RE-SENet is designed for the scenario of relation extraction task.This module is based on the SENet module,and in Squeeze operation,a global max pooling operation is added to increase the extraction of key features in the channel dimension;in Excitation operation,the Relu activation function in the bottleneck structure is replaced by the Mish activation function to reduce the semantic feature loss during the nonlinear fusion of channel features.Secondly,a pairwise ranking loss function CS-PRLoss that incorporates cost-sensitive learning is proposed for the problem of unbalanced relation categories commonly found in relation extraction datasets.CS-PRLoss introduces a cost-sensitive weight factor and the value is determined according to the number of samples of different categories,thus relation extraction model can focus more on learning the features of samples with less number of categories on the basis of effectively overcoming the influence of NR type samples in the training process.Thirdly,a relationship extraction model MC＿RS＿CP based on RE-SENet and CS-PRLoss is proposed.The model uses Bio BERT pre-trained language model as word embedding,uses Bi GRU network to obtain contextual feature information of text,uses multi-scale convolutional neural network to extract word combination features in text,and is further enhanced by the convolutional feature enhancement module RE-SENet,and uses the loss function CS-PRLoss to calculate the model loss.Finally,experimental validation and result analysis work is carried out on the DDIExtraction 2013 relation extraction dataset to verify the validity of the relevant research work in this paper.