Research On Aspect-based Sentiment Analysis Based On Gated Convolutional Neural Network

Aspect-based Sentiment Analysis(ABSA),as a fine-grained sentiment analysis method,can be used to determine the corresponding sentiment tendency of specific text targets.In order to solve the problem that the existing attention mechanism based ABSA methods pay more attention to the relative position features of words and ignore the absolute position features of context,a novel sentiment analysis model Pos ATT-GTRU-ABSA based on gated convolutional neural network is proposed.Firstly,the Laplacian kernel function was used to construct the relative position features.At the same time,the sentence length,the word position number and the size of the convolution kernel were taken as the absolute position features to construct the four-dimensional position matrix,and a new position processing method was proposed.Secondly,text location and text content are semantically encoded by using different sizes of convolution to obtain text convolution feature graphs and location convolution feature graphs.Then,Pos ATT-GTRU,a GTRU gated model with enhanced positional attention,is used to receive positional information,aspect information and content information.Attentional mechanism and two non-linear activation functions,relu and tanh,are used to control the semantic features,location features and aspect information flow to capture the dependency of long distance text.We get the attention eigenmatrix.Finally,the attention feature matrix is integrated into affective attention representation by using the full connection layer,and the final affective classification result is obtained by softmax function.Cross-entropy loss function is used to train the model,and accuracy is used as the main evaluation index.In this paper,data sets are used to verify the validity of the model,and the model is compared with ATAE-LSTM,Mem Net,IAN,RAM,GCAE,Mul-AT-CNN,PBAN,and IRAN.The experimental results show that the model in this paper achieves the optimal value of accuracy,which is 9.69%,8.90%,7.22%,5.10%,4.27%,3.79%,2.48%,2.13% on average,and38.07 times and 85.32 times of the average speed increase compared with the model in terms of running time.It can extract text position and word order features more effectively and realize the expression of local semantic features to deeper emotional features.There are 28 graphs,8 tables and 87 references in this paper.