| Since the 21 st century,the demand of information extraction has become more and more urgent.As an important task in text information extraction,relation extraction provides networked knowledge service for medicine,education,justice,safety,food and other fields,which plays a huge role in today’s society.This thesis mainly studies the relation extraction using deep learning method,proposes different solutions to the noise problem and Chinese corpus and constructs the corresponding network model.The main contents are as follows:A network model based on residual shrinkage structure is proposed to solve two kinds of noise problems in relation extraction.To solve the problem of noise inside sentences,a residual shrinkage network is constructed by adding soft thresholding structure to the residual network.The network utilizes the characteristics of deep learning to train different thresholds for different inputs and selectively delete some redundant features according to these thresholds.Experimental results show that this method can effectively solve the noise problem of single sentence in relation extraction.For the problem of noise between sentences in a packet caused by mislabeling under remote supervision,this study uses the attention mechanism to encode sentences in a packet and assigns different weights to each sentence,so that the sentences related to the current classification occupy a higher proportion.Experimental results show that these two noise processing methods can be used in parallel to reduce different kinds of noise and achieve better classification effect.To solve the problem of incomplete extraction of Chinese corpus,a fusion multi-word feature network model based on Bert and convolution is proposed.Firstly,the characteristics of Chinese and English are analyzed,and it is concluded that Chinese pays more attention to the expression of word semantics,while English pays more attention to the expression of sentence structure.More features at the lexical level were added to the model(including entity type,entity pair and keywords),which were sent into the model training together with the word vector of the original sentence.Here,the Bert model with strong feature extraction ability is selected to process sentences,and the convolution structure is used to process lexical features.Finally,the results of the two parts are fused and reclassified.Experiments show that introducing additional artificial features and using Bert and convolution combined network training can effectively improve the performance of Chinese relation extraction model,and obtain higher F1 values on the experimental data set. |
