Aspect Level Sentiment Analysis Based On Deep Learning

With the development of Internet technology,more and more people are willing to comment on online platforms.As time goes by,the network platform generates and saves a large amount of commentary text information.These reviews often contain users' tendentious views and sentiment on a certain event or a certain product,which is helpful to network public opinion monitoring.It also helps sellers improve product quality and service.In text sentiment analysis methods,coarse-grained sentence-level sentiment analysis can only obtain the overall sentiment of the text,but cannot obtain the sentiment information of different aspects or attributes of the object being reviewed.The fine-grained aspect-level sentiment analysis can obtain the sentiment information of each aspect or attribute in the review text,so that the public opinion,goods and services can be more accurately analyzed.This paper focuses on aspect-level sentiment analysis.The main research contents and contributions are as follows:(1)The existing aspect level sentiment analysis seldom considers the syntactic relationship between words,which may lead to the wrong matching of aspect words and sentiment words and reduce the performance of aspect level sentiment analysis.Therefore,an aspect level sentiment analysis model based on distance and graph convolution network is proposed.Firstly,a bidirectional LSTM network with residual connection is designed to extract sentence context;secondly,a syntactic distance weight module is designed,which can obtain the syntactic distance weight of words according to the syntactic dependency tree,and construct the adjacency matrix according to the syntactic relationship between words;finally,the graph convolution network is used to extract sentence context,syntactic distance weight and syntactic adjacency matrix Based on the feature of sentiment,aspect level sentiment analysis is realized.Experimental results show that on the Twitter,Laptop14 and Restaurant14 datasets,compared with the existing models,the accuracy of the proposed model is1.4%,1.9% and 1.7% higher than that of the existing model respectively,and the MF1 values are 1.5%,1.3% and 1.6% higher than of the existing model respectively.(2)This investigating is based on multi task learning and aspect level affective analysis of BERT.Most existing aspect-level sentiment analysis ignore the implicit expression of aspect categories,sentiment analysis when no obvious aspect words are given in the sentence.Therefore,this paper proposes an aspect level sentiment analysis method based on multi task learning and Bert.Firstly,the sentence and the preset aspect category are input into the BERT model to obtain the sentence representation of the fused aspect category,and the weight of the aspect category in the sentence is calculated.Secondly,use the sentence representation output by BERT as the input of the fully connected layer,calculate the sentiment probability distribution of each word.Finally,the sentiment probability distribution of each word is multiplied by the weight of each aspect category to obtain the emotion of all preset aspect categories,to realize aspect-level sentiment analysis.Experimental results show that on the Restaurant14 and MAMS datasets,the accuracy is 2.5% and 4.7%higher than of the existing model respectively,and the MF1 values are 0.9%,and5.5% higher than of the existing model respectively.