Research On Text Sentiment Analysis Method Based On Adversarial Training And Multi-layer Attention Mechanism

As a fine-grained sentiment analysis task,aspect-based sentiment analysis aims to predict the emotional polarity of different aspects in a sentence.In recent years,with the rapid development of network technology and social media,it is still a hot research to complete the task of aspect-based sentiment analysis by means of deep learning.In order to achieve the purpose of analyzing different aspects of emotional orientation,the fine-grained information implied between aspect words and context words is considered.Firstly,a hybrid neural network aspect-based sentiment analysis based on MHSA(HCB-MHSA)is proposed,which gives full play to the ability of CNN to extract phrase-level features of text and Bi LSTM to extract global structure information of text,so as to achieve the goal of fully mining the implied semantic features of text.At the same time,the MHSA is introduced to capture the overall sentiment characteristics of the sentence,effectively solve the problem that the information in a long distance is weakened,and accurately extract the relevant information.In addition,the position embedding module is introduced to enhance the word representation in sentences.Secondly,based on the HCB-MHSA model,because a single attention mechanism cannot provide accurate weight scores for the judgment of the emotional polarity of the model,and the neural network method cannot effectively obtain the hidden representation of sentences,an aspect-based sentiment analysis method(MASEGCN)using BERT and multi attention mechanism is proposed.Consider using use the MHSA to obtain the weighted attention scores of context words and aspect words,and then use the interactive attention mechanism to model aspect words and context interaction.In addition,a feature enhanced attention mechanism is proposed to improve the accuracy of sentiment analysis.Finally,the existing research shows that applying GCN to aspect-based sentiment analysis has a good effect,so based on the MASEGCN model,considered to input the feature vectors of sentences and context words after syntax and grammar enhancement to GCN for feature extraction to obtain dependency and semantic information between words.To solve the problem that the labeled dataset in aspect-based sentiment analysis is too small,considering building a specific BERT based on adversarial training,and obtain the current gradient by accessing the model parameters to generate adversarial text,so that the size of the dataset can be expanded,and effectively improve the classification accuracy of the mode.Finally,the experimental results on four datasets Lap14,Rest14,Rest15 and Rest16 show that our proposed model has achieved better results than many current baseline models.