Aspect-level Sentiment Analysis Based On Global And Local Features Of Orthogonal Projection

Aspect-level sentiment analysis aims to identify the sentiment information expressed by various aspects of a sentence(e.g.quality,price)and to determine the sentiment relationship between the expression of a word or phrase in an aspect and the correct description of the sentiment of a word or phrase in an aspect.This research task can provide theoretical support for information extraction,opinion mining,intelligent recommendation,public opinion monitoring and other applications.The existing deep learning methods focus more on mining local information,and there is a problem of weak global dependence information mining.An aspect-level sentiment analysis method based on global features of orthogonal projection and local dependence fusion graph convolution is proposed.The main research contents are as follows:(1)A Bi-LSTM-CNN sentiment feature extraction model based on orthogonal projection is proposed to mine deep semantic sentiment features,obtain sentiment features with high discrimination,and solve the problem of limited semantic information for feature extraction.Firstly,the neutral word vector is projected into the orthogonal space of sentiment polarity words to obtain the weighted neutral word vector,and the key semantics of the text are extracted through the CNN deep learning model.Then,using the Bi-LSTM-Attention model and the weighted neutral word vector,learning from the extracted key semantics can enhance the semantic features of sentence sentiment and make the text more discriminative in sentiment classification.Experimental results show that the proposed model significantly improves the accuracy of text sentiment classification,which is sufficient to illustrate the effectiveness of the proposed sentiment feature extraction method.(2)A convolution method of global feature and local dependency fusion graph based on orthogonal projection is proposed and applied to aspect-level sentiment analysis to mine more complete text sentiment features and solve the problem of weak mining of global and local feature information.First,simplify the simplified global feature structure diagram of the text.BERT is used for pre-training,and the sentiment feature extraction method of orthogonal projection effectively weakens the dependence of node updating in the graph and reduces the dependence between nodes and the whole corpus.This method retains the global feature information and significantly reduces the number of edges and memory consumption.Secondly,the local dependency information of sentences is mined by using syntactic dependency structure and sentence sequence information.Then,a percentile multiread attention mechanism is used to measure key parts of the GCN output for better sentence representation of a given aspect.Finally,position coding is introduced to simulate the specific representation between each aspect and its context,which effectively combines the global and local dependency structure features.The experimental results show that the proposed global feature and local dependency fusion graph convolution method based on orthogonal projection can not only effectively mine the feature information hidden in the database,but also effectively improve the accuracy of the aspect-level sentiment analysis of the review text.Figure [21] Table [9] Reference [73]...