Research On Text Sentiment Analysis Method Based On Hybrid Attention Mechanism

The rapid development of the network has improved people’s living standards,and at the same time,there have been many problems to be solved,including sentiment analysis.Sentiment analysis can be applied to various realistic scenarios such as commodity review analysis,public opinion analysis,and film review analysis.Mastering efficient can often help decision makers to formulate more reasonable solutions and avoid wasting time and resources.For the problem of sentiment analysis,researchers have proposed a variety of solutions,but not many are dedicated to the Chinese field,and most of the existing models have the problems of huge amount of model parameters,slow training and low accuracy.In order to solve these problems,this thesis will take aspect-level sentiment analysis as the starting point and complete the following work:(1)The BERT(Bidirectional Encoder Representations from Transformers)algorithm has been optimized.BERT cannot extract word vector features when processing Chinese text.For this reason,AWC(Average Word Vector Convolution Module)is proposed.By introducing an attention mechanism into the convolutional neural network to extract reliable word vector features,and then further obtain the local features of the text,which makes up for the shortcomings of BERT’s inability to extract word vectors.At the same time,the balanced multi-head design,hierarchical parameter sharing mechanism,convolutional connection are used to optimize the model structure,which greatly reduces the amount of model parameters,and finally forms a coarse-grained sentiment analysis algorithm based on the hybrid attention mechanism.(2)In order to strengthen the correlation between aspect words and context and text words,this thesis uses the optimized BERT model as the word vector encoding module to obtain context,individual words and aspect word information.At the same time,the information interaction module IEM(Information Interaction Module)is designed in combination with the attention mechanism.This module can obtain the relevant information between the aspect word and the context,the aspect word and the individual vocabulary,so that the aspect feature and the context feature can be efficiently integrated,and finally realized by the softmax function.(3)Fine-grained sentiment analysis is a current hot research topic,but there is a lack of relevant Chinese data.This thesis uses crawler as the main method to obtain online shopping review data,and combines the computer data set format proposed by the International Semantic Evaluation Conference to clean the crawled data,and finally obtain a Chinese aspect-level sentiment analysis data set.Through experimental comparison,the two algorithms proposed in this thesis are better than the comparison algorithms in processing Chinese data,and the model parameters are less,model design achieves expected results.